Boxing is widely recognized as a physically demanding and inherently
dangerous sport. The World Medical Association has raised particular
concern, stating that the “main medical argument against boxing is the risk
of chronic traumatic encephalopathy (CTE)”, a degenerative brain condition
linked to repeated head trauma. Beyond long-term neurological risks, acute
injuries are also prevalent in competitive boxing (G., Magdalena, Lindsay,
I., & M., 2019).
A comparative study analysing 7,712 athlete exposures across three
consecutive Olympic Games emphasised the pressing need for “injury
prevention initiatives to reduce the burden of injury among combat sport
athletes.” Notably, the findings revealed that injury risk in boxing shows
“no difference by sex, weight category, or tournament round,” accenting that
the hazards of the sport affect athletes indiscriminately, regardless of
demographic or competitive context (Lystad, Alevras, Rudy, Soligard, &
Engebretsen, 2021).
Given these challenges, there is a growing necessity for solutions that
mitigate injury risk during training. This is especially crucial for elite
athletes, where sparring remains an indispensable component of preparation
but simultaneously exposes them to the highest risk of harm. Therefore,
alternative training methods are needed approaches that replicate the
intensity and realism of sparring while offering a safer environment.
2. Problem Clarification (Zakir)
2.1 Background
Sparring is widely regarded as the cornerstone of boxing training, as it
represents the “combat” element of the sport and enables athletes to put
their techniques into practice under realistic conditions. It is during
sparring training that boxers sharpen essential skills such as timing,
defence, and adaptability under pressure. However, research indicates that
“repetitive sub-concussive trauma from frequent and intense sparring
training appears to be a significant risk for developing neurological
sequelae”, highlighting the health risks associated with its practice
(Stiller, et al., 2014). While sparring training remains indispensable,
these risks point to the need for safer alternatives that can replicate its
training benefits without compromising athlete well-being.
To better understand this problem, our team engaged directly with the boxing
community by signing up for boxing gym memberships. We conducted user
interviews, alongside a secondary survey of existing solutions and
competitors in the boxing training landscape. From these findings, we
developed two value proposition canvas to capture the needs and pain points
of both boxers and the community.
2.2 Primary Research: User Interviews/Surveys
2.2.1 Methodology
Our research team conducted user interviews at six different boxing gyms in
Singapore to broaden the scope of perspectives.
Our research method was guided by the Interaction Design Foundation (IDF)
framework, where user interviews are carried out during the empathize phase
of the design thinking process (IxDF, 2016). While at the various boxing
gyms, we employed IDF’s method of Unstructured Interviews, enabling a
“qualitative exploration of users’ thoughts and experiences” (IxDF, 2016).
In addition, we also designed and sent out survey forms to get more
viewpoints from boxers that we did not meet in-person from the
aforementioned gyms. Through our user interview and surveys, we defined the
different stakeholders in the boxing community into two groups;
- Boxing-Gym Boxers: People who train recreationally or
competitively in boxing gyms are known as boxers.
- Training Support Stakeholders: Stakeholders in the
training support ecosystem include parents, trainers, and gym owners who
plan, coordinate, and encourage boxing involvement and progress.
Questions are tailored accordingly and refined them into a concise set of
guiding questions (Table 1).
2.2.2 Insights
Our user interview and surveys of Boxing-Gym Boxers presented that sparring
training still accounts for about 22.6% + 15.1% (≈37.7%) of reported
difficulties highlights the fact that sparring training continues to be
boxers' main source of both attraction and annoyance (Figure 1). Finding
reliable partners who can match their ability and energy levels is a
challenge for many boxers, which can result in lower-quality training or a
higher risk of injury. However, about 20.8% of respondents say they would
like to learn and advance more quickly but feel constrained by the format of
the lessons that are offered. Because different skill levels are frequently
mixed in group training, more experienced boxers are forced to hold back or
repeat fundamental drills, which hinders their progress and saps their
desire. Essentially, even while sparring training is essential for
development, the way gyms are set up now frequently makes it difficult to
customize the level of intensity and learning rate, which leaves experienced
boxers constrained in development and those looking to improve constrained
by inflexible class schedules.
Approximately 33% of coaches within the ‘Training Support Stakeholders’
group prioritized pad work, and 22% thought sparring training sessions were
the best way to keep boxers engaged (Figure 2). These results imply that
sustaining motivation and skill improvement requires direct, coach-like
engagement, whether via guided drills, pad work, or sparring. However, with
current boxing gym setups, such individualised attention is not possible as
the coach to boxer ratio is uneven.
As the basis for the Value Proposition Canvas, we provide insight summaries
for Boxers at Boxing Gyms and Training Support Stakeholders, emphasizing
their primary tasks, challenges, and rewards.
These realizations served as the basis for the Value Proposition Canvas's
Customer Profile. Each group stands for a distinct set of jobs, pains, and
gains made during the interview. Focused value propositions and product
features are then created that support their objectives by comprehending
their unique motivations and frustrations, as revealed in later parts.
2.3 Secondary Research: Competitor Analysis
Following the identification of user insights, a competitor study was
carried out to look at current boxing training programs and spot any
possible market gaps. The goal of this analysis was to determine how
existing goods either meet or fall short of meeting the pains, gains and
jobs identified by our user interview. We determined what people value most
and where discontent frequently arises by analysing the features,
functioning, and user feedback of each competition. By evaluating each
competitor’s features and online reviews, we can identify what users value
most and where discontent commonly arises.
Three close competitors were chosen for a thorough comparison in order to
guarantee relevancy; each was picked because of their high internet
presence, which indicated a sizeable percentage of user adoption (Table 3).
Furthermore, our competitor selection intentionally focuses on ‘robotic’
training equipment as the primary problem identified in our research is the
lack of adaptive training partners. These contrasts gave us a standard by
which to measure our own value proposition and design approach.
Our innovation strategy was guided by the useful design orientations and
functional gaps that were discovered through the examination of features
and feedback across a number of commercially available solutions (Table 4).
On-screen coaching, variety of strike zones, impact sensors, and a range of
training modes (e.g., cardio, sparring, mitts) are already features of existing
systems, such as interactive boxing trainers and AI-powered punch trackers.
These features suited to various training preferences, offered quantifiable
performance statistics, and successfully increased user motivation. Nevertheless,
significant drawbacks remained, such as unreliable technology, no reactive or
tactile feedback, irrational attack angles, and an inability to simulate defensive
situations. Furthermore, a lot of sophisticated technologies were either too
expensive, lacked adaptive intelligence, or were too immobile to accommodate dynamic
boxing motions like footwork.
Analysing competitors revealed that although existing solutions effectively use
digital coaching and quantifiable data, they frequently fail to provide realistic
engagement, adaptability, and safety for a wide range of users. Many technologies are
still too expensive for general use or too inflexible for practical training. Our team
used these insights to create five integrated products that each addressed a distinct
market gap: Intelligent Training System, Skill Progression Studio, Adaptive Fight
Intelligence, Performance Analytics, and Modular Boxing Platform. When taken as a whole,
these solutions transform boxing training by fusing accessibility, realism, and data
-driven feedback to benefit gym-based stakeholders as well as individual athletes.
2.4 Value Proposition Canvas
Two Value Proposition Canvases were created to reflect the unique but related demands of
Boxing-Gym Boxers (Figure 3) and Training Support Stakeholders (Figure 4) after-market gaps
and user insights were identified. These canvases convert the recurrent themes, frustrations,
and intended results from our secondary research and interviews into focused product
opportunities.
For the Boxing Gym Boxers, the findings showed that boxers had a high demand for
flexibility in training, quantifiable development, and realism. Many complained about the uneven
quality of the sparring-drill sessions, the lack of feedback in class, and the challenge of
striking a balance between safety and growth. This influenced the development of solutions that
give controlled and flexible skill development, duplicate the intensity of sparring-drill sessions
without the risks involved, and provide measurable feedback to maintain motivation and progress.
Safety, structure, and scalability were the main concerns for Training Support Stakeholders.
Interviews brought to light issues such a lack of gym space, expensive operating expenses, and the
difficulties of sustaining participation across a range of ability levels. In order to reassure
parents and boxers about the efficacy of training, stakeholders looked for methods that enhance
safety oversight, standardize the quality of instruction, and offer quantifiable performance data.
The Intelligent Training System, Skill Progression Studio, Adaptive Fight Intelligence, Performance
Analytics, and Modular Boxing Platform are five interconnected solutions that were designed with the
use of these insights. Each addressing the challenges and successes noted in both canvases, striking
a balance between the stakeholders' demands for safety, structure, and operational effectiveness and
the boxers' desire for realism and advancement.
2.4.1 Problem of Interest
A basic problem that both boxers and training support stakeholders face is the absence of a thorough,
secure, and flexible training program that combines the structure and quantifiability of contemporary
coaching with the reality of sparring-drill sessions, according to the two Value Proposition Canvases.
While boxers want flexible training, realistic involvement, and regular feedback, coaches and gym owners
struggle to provide safety, personalized instruction, and efficiency in a constrained amount of time and
space. Only specific elements of the training experience are addressed by current training technology.
None offer a single platform that supports teaching, data tracking, and realistic skill development all
at once, allowing for safe, intelligent training. This restriction causes a disconnect between what
coaches can provide during training sessions and what boxers want for realistic practice.
Only specific elements of the training experience are addressed by current training technology.
None offer a single platform that supports teaching, data tracking, and realistic skill development all
at once, allowing for safe, intelligent training. This restriction causes a disconnect between what coaches
can provide during training sessions and what boxers want for realistic practice. Therefore, our goal is to investigate
"How might we design an intelligent boxing training system that can dynamically and safely strike with users while
assisting coaches in guiding, monitoring, and sustaining boxer development both during and beyond training sessions."
This serves as the basis for our design direction, which aims to develop a human-assistive training ecosystem that
redefines how boxing skills are learnt, practiced, and mastered by serving as a training partner for athletes and an
educational extension for coaches.
2.4.2 Value Proposition Statement
Based on insights from interviews, surveys, and analysis of existing solutions,
the proposed robot directly addresses challenges faced by boxers, including
lack of training variety, misalignment with body types, and limited ways to
track progress. At the same time, it enhances motivation and inclusivity by
introducing interactivity and personalized feedback. Thus, our value proposition
statement is defined as:
“The boxing robot delivers a safe and dynamic training experience that adapts to
each individual, offering the realism of training with a training partner, and
the guidance of a coach while supporting users of all levels to train effectively,
build confidence, and continuously improve.”
3. Design Methodology (Zakir)
To translate the problem of interest and value propositions into actionable design
outcomes, this project adopted Pahl and Beitz (Pahl, Beitz, Feldhusen, & Grote, 2007)
approach to engineering design. Their framework emphasizes a logical, iterative progression
from problem definition to detailed realization, ensuring that all design decisions remain
traceable to user needs, competitor insights, and technical requirements.
According to Pahl and Beitz (Pahl, Beitz, Feldhusen, & Grote, 2007), there are four main stages of design:
Clarification of Task: Involves understanding and defining the design problem, identifying user and technical requirements, and outlining key system functions.
Conceptual Design: Focuses on developing alternative principles and mechanisms that could satisfy these functions.
Embodiment Design: Refines selected concepts, assessing feasibility, structure, and integration across subsystems.
Detail Design: Finalizes the chosen concept for prototyping and testing.
Each phase is supported by specific design tools developed for this project:
Clarification of Task → Product Needs Translation
Conceptual Design → Solution Principles and Concept Generation
Embodiment Design → Concept Screening
Detail Design → Prototyping and Testing
Every choice, from determining market gaps to improving technical performance, was supported
by methodical reasoning through this organized design technique. It offered the engineering
expertise required to turn research findings into a logical, useful, and user-focused boxing
training program.
3.1 Clarification of Task
Insights from the competition study were first translated into specific product requirements
using the Product Needs Translation tool (Table 5), which identified important enhancements,
performance standards, and user experience objectives. The Function Specifications, which included
explicit functions, performance standards, and design justifications for the robot's subsystems
(arms, head, legs, body, skeleton, and interface), were subsequently influenced by these translated
needs.
4. Final Design (Jeanette and Zakir)
4.1 System Overview (Jeanette)
Building on the identified product needs, the BoxBunny robot is organised
into four key subsystems, each addressing a distinct set of functions
(Figure 5):
User Interface – touchscreen, visual feedback, and
control logic for configuring and running training sessions.
Upper Mechanism – includes Punch Execution (arms),
Contact Points (head/torso/arms), and Impact Sensors responsible
for safe, realistic striking interaction.
Lower Mechanism – comprises Footwork Movement
(legs), Compact Footprint (base), and Height Adjustability, which
together position the robot correctly relative to the user.
Robot Intelligence – primarily User Tracking,
enabling the robot to respond to the boxer’s movement and adapt
drills in real time.
Each of these subsystems is detailed in the following sections.
4.2 User Journey (Zakir)
4.2.1 End-to-End User Flow
The flowchart in Figure 6 illustrates the complete user journey when interacting with the robot,
from the start of a session to its conclusion. Each activity is color-coded according to its
corresponding product feature category, as indicated in the legend on the right. The session begins
with height adjustment and selecting the training parameters that match the user’s goals. Once training
starts, the robot functions as an intelligent training partner. The robot enables users to strike designated
zones, defend against simulated punches through arm actuation, and move or orient itself in response to the
user. At the end of the session, the robot generates performance analytics displayed on the dashboard. From
there, users can choose to begin a new session or exit, depending on their needs.
4.2.2 Training Modes Overview
The robot offers two primary training modes: Training and Performance.
Firstly, Training Mode includes Techniques and Spar.
Spar features five preset boxer types that simulate common opponent styles.
Techniques provides three categories: Punch Combination Library, Offensive Drills, and Defensive Drills.
The Punch Combination Library includes multiple difficulty levels and also allows users to customise their own punch sequences.
o Offensive and Defensive Drills focus on developing the user’s specific attacking and defensive skills respectively.
Performance Mode measures key physical attributes such as stamina, punch power, and reaction time.
Additionally, the Others menu on the homepage enables users to adjust system settings such as switching
between orthodox and southpaw, while the Insights feature provides a summary of training metrics and performance
trends.
4.2.3 Movement Behavior
When the user selects the training parameters, they are interacting with the Skill Progression component of the
product. This feature determines the appropriate motor actuations and software activations (such as computer vision)
required for the session. As shown in Figure 7, users can choose from several training categories such as Training,
Performance, and Others with each serving different purposes. For clarity, this report focuses only on the green pathway,
as detailing every iteration would be unnecessarily exhaustive.
The green pathway leads to the Counter Punch Mode, which is intentionally chosen because it engages the full range of the
robot’s capabilities. This mode is for all skill levels, as it replicates one of the most common foundational sparring
opponents: a reactive boxer who waits for the user to initiate an attack and responds with precise, well-timed counters
(typically jabs and crosses). This style is characterized by low aggression, high timing sensitivity, and reactive movement.
Translated into robot behaviour, the system uses computer vision to detect the user’s punches. For every punch thrown by the
user, the robot responds with a counter-punch by actuating its arm motors. Additionally, the robot adjusts its position by
moving linearly when the user steps forward or backward, and it maintains proper orientation by rotating to face the user.
This ensures a realistic and continuous sparring interaction.
The Graphic User Interface (GUI) fulfils the product needs of Skill Progression and Performance Analytics by providing a
modular, intuitive training environment that guides users through structured development pathways. It enables BoxBunny to
select and execute progressive training modules from offensive and defensive drills to measuring performance. The interface
organizes these programs into clear categories, ensuring balanced skill advancement across technique, sparring, and performance
tasks. Additionally, the GUI serves as a comprehensive dashboard, visualizing user statistics across sessions to motivate
improvement, support self-assessment, and foster healthy competition. Users can conveniently select training parameters, as
shown in the demonstration below, making the GUI the central touchpoint for both training customization and performance insight.
5.1.1 Function Specifications
The User Interface subsystem processes user input and delivers visual feedback with responsiveness suitable for real-time training
interaction. To ensure clarity, comfort, and seamless user control, the system must meet the following performance targets:
Meeting these specifications ensures that the interface remains readable in various lighting conditions and responds quickly
enough for real-time adjustments during training. This makes the User Interface subsystem a reliable control and feedback channel,
supporting uninterrupted immersion and natural user–robot interaction.
5.1.2 Design Considerations
The principal design consideration for the screen-based user interface is achieving a reasonable resolution while ensuring the
display remains easy to integrate into the overall system. The chosen display must provide sufficient clarity for users to read the
screen without visual strain or complicate hardware compatibility. Equally important is seamless integration with the robot’s
electronics and software stack; any display that requires complex interfacing, uncommon drivers, or extensive modification introduces
unnecessary points of failure. Consequently, the primary design challenge lies in selecting a display that balances visual clarity with
practical system integration, ensuring reliable performance without imposing additional hardware or software overhead.
5.1.3 Concept Generation
Concept 1: Developing a Mobile Application
The first concept explores the idea of using the user’s personal smartphone as the primary interface. This approach promises high-resolution
visuals at minimal hardware cost to the robot. However, the development effort required to build and synchronise a mobile app with the robot’s
embedded electronics was deemed excessively complex. In addition, this solution introduced significant usability constraints: in a gym environment,
users would need to rotate their personal phones when taking turns, and coaches might not have their devices readily accessible during training.
In some cases, facilities might even need to purchase a dedicated phone solely for operating the robot. These practical limitations made the concept
unsuitable for reliable day-to-day use.
Concept 2 (Selected Concept): A Dedicated Screen
The second concept proposed integrating a dedicated onboard display. This screen could interface directly with the robot’s electronics, greatly
simplifying programming, communication, and system reliability. It also provides a consistent user experience without dependence on personal devices.
However, this approach increases hardware cost and typically offers lower resolution compared to modern smartphones, as an affordable display must be
chosen to keep overall system pricing reasonable. Despite these trade-offs, the dedicated screen avoids the operational and integration challenges
present in Concept 1, making it the more practical solution.
5.2 Upper Mechanism (Elgin)
The Upper Mechanism constitutes a primary subsystem of the robotic boxing
trainer. It is designed to simulate the upper-body actions of a training
partner, encompassing punch executions and the necessary contact points.
This mechanism is comprised of two key components: the "Punch Execution"
system, and the "Contact Points".
For the development of an effective training robot, an initial analysis of
traditional boxing training methods was conducted. The primary method,
"padwork," involves a coach or partner holding pads (or "mitts") for the
boxer. Padwork is highly effective for training accuracy, technique, and
combinations. Its effectiveness, however, is highly dependent on how the
pads are held. A skilled coach holds the mitts close together to simulate
the compact frame of an opponent’s head or torso, compelling the boxer to
refine their technique and mechanics, focusing their power effectively.
However, should the training partner be inexperienced, they might fail to
provide sufficient resistance or catch the punch improperly. This can result
in over-extension injuries for the boxer, such as hyperextension of the
elbow or significant shoulder strain.
Flexible training sticks are soft, flexible, lightweight, and a low-cost
alternative to traditional pads. This design almost entirely mitigates the
fear of injury from impact, both for the boxer and the coach. For defensive
drills, the coach swing the noodles to simulate punches. In
the absence of fear, the boxer can engage in more realistic scenarios,
practicing slips, bobs, and weaves against a moving, non-threatening object.
This is ideal for training reactions, speed, and technique. The minimal
profile of the sticks also maintains the accuracy-training benefit of pads.
From this foundational research, the general design criteria for the robotic
mechanism were established which leads into the specification for section.
5.2.1.1. Design Criteria
The robot must provide realistic, anatomically correct striking spots for
the boxer, specifically the head, celiac plexus, and liver.
Striking Zones
Target Motion
The following target motions, derived from coaching techniques, were used to
guide the robotic arm's design. The system must be able to replicate these
three distinct strikes.
Based on the preceding research, the following performance metrics were
established through physical measurements and validations.
5.2.2.2.Design Considerations
5.2.2.3. Final Concept
The final design selection was a 2-Degrees-of-Freedom (2DOF) mechanism,
which was determined to offer the optimal balance of fidelity, complexity,
and cost by utilizing only two motors per arm. This 2DOF system is comprised
of two primary axes. The first is a pitch axis, which provides vertical
rotation and permits the execution of the standard downward jab. The
inclusion of a secondary yaw axis, providing horizontal rotation, enables
the system to execute both the horizontal hook and the diagonal uppercut.
Through the coordinated combination of these pitch and yaw axes, this single
compact mechanism can be programmed to create realistic, and fluid strike
patterns, successfully simulating all required training motions. To mitigate
inertia and motor load, the training sticks will be constructed from
polyethylene foam. This material is selected for its low density and high
resilience, a choice that substantially reduces the system's moment of
inertia. A functional prototype has been 3D printed and has been used to
demonstrate the various striking requirements.
A crucial function specification pertains to the user's physical safety
relative to the machine. To ensure a user does not get too close and
inadvertently trip over the robot's base, a key spatial metric was
established. This positional constraint dictates that the striking pads must
be positioned 40 cm from the centre of the robot.
5.2.3.2. Design Considerations
The padding serves as the primary target for the user, and its design is
governed by three key considerations. First is fidelity; the padding must
represent the key striking areas of the head, liver, and celiac plexus.
Their placement must be anatomically correct to simulate a real opponent.
Second, the delimited size of these areas is a deliberate design choice to
foster accuracy, ensuring the boxer is training for technique by targeting
small, specific zones rather than a large, imprecise area. Finally, for the
safety of both the boxer and the robot's internal components, the padding
must possess a soft contact surface and be capable of absorbing the strike
energy from average users.
5.2.3.3. Solution Principles
This implementation will be achieved by first utilizing a standard
multi-layer padding structure, consisting of a soft contact layer for user
comfort and a dense, energy-absorbing core for force dissipation. Second,
anti-vibration rubber mounts will be employed to attach the padding assembly
to the robot's main frame. This isolation is critical for protecting the
robot's sensitive internal electronics from the repeated impacts. Finally,
the integration of springs may be considered in addition to the mounts,
providing additional dampening.
The primary objective of this component is to detect a successful punch
landed on a padding-integrated contact point. This provides immediate,
positive feedback, which is crucial for user motivation and the motivational
aspects of the training session. Furthermore, this data is slated for
integration into training analytics, which will provide a post-session
report to the user tracking metrics such as punch count, speed, and reaction
time. Therefore, the sensor data must demonstrate high consistency and
reliability, even during active robot use.
The principal design consideration for the user input system is sensor
durability. The selected sensors will be subjected to high impacts, which
represents the most critical potential failure point for this subsystem.
Consequently, the primary design challenge is the mitigation of wear and
tear, especially that which is caused by direct load or force.
5.2.4.2. Solution Principles
During concept generation, initial considerations focused on direct-load
sensors, such as integrated buttons or Force Sensitive Resistors (FSRs).
Although these components are low-cost and simple to implement, they suffer
from a critical flaw: they are placed directly in the line of force. This
positioning means they are subjected to the full, direct load during every
impact, which can lead to physical degradation.
An alternative involves indirect-load sensors, which are not positioned in
the direct line of force. One such approach is to use an accelerometer,
which can be mounted on to the padding assembly. This sensor would be
configured to detect the sharp vibrations characteristic of a punch. Another
indirect method is an Infrared (IR) proximity sensor, which could be
recessed within the pad holder and configured to detect the movement of the
pads during an impact. The primary advantage of both solutions is
significantly longer durability. However, these sensors introduce a new
challenge: signal noise. The data handling will therefore require robust
signal processing and filtering to accurately distinguish a punch.
5.3 Lower Mechanism (Jeanette)
The lower mechanism integrates three functions: Footwork Movement, Compact
Footprint, and a Height‑Adjustment Interface, to align the robot’s position
with the boxer’s stance and approach during training.
The design intent is to:
Enable structured, realistic footwork movement in drills that
develop ring generalship (control of distance, timing, and angles).
Maintain a compact, safe physical footprint that minimises
collision and trip hazards within crowded gym environments.
Provide sufficient vertical adaptability so that head and torso
targets can be matched to different user physiques in a fail-safe
manner.
This section presents the functional context and specifications of the lower
mechanism, the associated design considerations, the key outcomes from
concept evaluation (detailed in Appendix B), and the final selected concepts
for the footwork, base, and height-adjustment subsystems.
In live padwork between coach and boxer, the coach continuously manipulates
both range and angle to shape the rhythm and difficulty of drills. These
adjustments are central to what is referred to as ring generalship, the
boxer’s ability to control pace, distance, and the flow of a bout through
strategic footwork and positioning. A boxer with strong ring generalship can
dictate the terms of engagement, forcing the opponent to step into traps,
overreach, or drift into disadvantageous positions (Evolve MMA, 2021).
Padwork observations and self-experiments highlighted several recurring
patterns:
Offensive drills: The coach steps back or out as
the boxer steps in with jabs and combinations, forcing the boxer to
close distance in a controlled manner.
Defensive drills: The coach pivots or re-angles to
present new attack lines, prompting the boxer to adjust footwork,
guard, and counter-punch timing.
Counter drills: After an initial exchange, the
coach rotates to open new angles, and the boxer is expected to
respond with appropriate foot placement and counters.
Footwork Movement - Step in Jab
Footwork Movement - Pivot Counter
From a mechanical perspective, these behaviours can be decomposed into two
essential motions:
Linear translation (approach and retreat) along a
primary training axis.
Yaw rotation (on-the-spot turning) around a
vertical axis to re-present different lines of attack and defence.
To replicate this in BoxBunny, the lower mechanism must provide controlled
fore–aft translation and yaw rotation, at
speeds and ranges that are consistent with padwork, while remaining stable
under repeated punching impacts.
5.3.1.2. Function Specifications
Translating the padwork behaviour into engineering requirements, the
footwork subsystem is specified by the target metrics in Table 7. Key design
considerations used as constraints during concept selection are listed in
Table 8. These were derived from self-experiments and safety considerations,
ensuring that robot motion remains realistic and responsive but not
hazardous.
These specifications and requirements were used as design constraints and
evaluation criteria for the concept generation and selection process
presented in Appendix B.
5.3.1.3. Motion Axis Selection
To implement the two required degrees of freedom (translation and yaw), two
main architectures were evaluated:
A single system based on
omnidirectional wheels
, where a common set of wheels provides X–Y translation and yaw
rotation.
A decoupled-axis architecture, where yaw rotation
and linear translation are handled by two separate mechanisms: a
slewing base for yaw and a linear stage for range.
The omnidirectional wheel concept is attractive in general mobile robotics
because it provides full planar mobility and compact actuation, and it
supports high-level commands such as “move forward”, or “rotate on the spot”
with a single hardware stack. However, as detailed in Appendix 2)a)i), it
presents several drawbacks for a punch‑receiving boxing robot due to its
dependence on floor conditions, limited positional stiffness, and coupled
motions. These issues are reflected in the decision matrices in Appendix B
(Tables A9 and A10), where a decoupled-axis architecture
using a slewing base at the bottom to provide yaw rotation
and an independent linear sliding stage above it to provide
forward–backward translation, is selected for BoxBunny’s footwork movement.
The evaluation confirms that the decoupled architecture outperforms
omnidirectional wheels in the high‑priority criteria of stiffness,
repeatability, and impact robustness. Accordingly, near the bottom of the
robot (Figure 12a), BoxBunny adopts
separate Slewing-Base Yaw and Linear Slide Range axes
(Figure 12b).
The key selection criteria for both axes are: (i) ability to hold position
under impact (angle and position rigidity), (ii) ability to carry combined
static and dynamic loads of the full robot and user strikes, and (iii)
provision of moderate speed with high torque for confident starts, stops,
and direction changes.
5.3.1.4. Linear Motion
For the translational axis, the final concept is a
motorised sliding-table linear rail driven by a ball screw
(Figure 13). This decision followed the evaluation of three power
transmission options (lead screw, ball screw, and belt drive), summarised in
Appendix (Power Transmission Drive Selection, Table A11).
The selected mechanism comprises:
A sliding table that serves as the platform for
the entire upper portion of the robot.
Two linear rails with preloaded blocks, which
guide the table and carry longitudinal forces and overturning
moments arising from both motion and punching.
A ball screw drive, which converts motor rotation
into linear motion of the sliding table.
The decision matrix in Appendix (Table A11) shows that the ball
screw–driven linear rail scores highest on the high-priority criteria of
load and thrust capacity,
positional stiffness
, and accuracy/repeatability, while remaining acceptable in
cost and maintenance. Therefore, a
ball screw–driven sliding table on linear rails
is adopted as the linear footwork mechanism.
5.3.1.5. Rotation Motion
For yaw rotation, the selected concept is a
slewing ring bearing with external gear teeth
, driven by a pinion gear and BLDC servo motor (Figure 14).
This solution emerged from the bearing and motor selection processes
described in Appendix B (Bearing Selection, Table B13; Motor Selection,
Table B14).
A four-point contact ball slewing ring with external gear
was selected as the yaw bearing because it can handle combined axial, radial
and overturning loads with good stiffness, while being cheaper and more
available than cross-roller types. A
BLDC servo motor with encoder
was chosen to drive this stage, as it offers low backlash, smooth torque,
precise position control and strong disturbance rejection under punching.
Having met the fundamental requirement of executing padwork-inspired
footwork motions, the next consideration is
safety and spatial efficiency
. Boxing gyms are often space-constrained, with multiple users training in
parallel. It is therefore essential that BoxBunny’s base has requirements
summarised in Table 9.
5.3.2.2. Base Design
To satisfy these requirements, the base is designed as a
trapezoidal platform
rather than a large circular or rectangular footprint (Figure 15). This
geometry is chosen for the following reasons:
The narrow front edge of the trapezoid reduces
obstruction near the boxer’s lead foot, leaving space for pivots
and lateral adjustments.
The wider rear section provides more area for
mounting the lower mechanisms and optional ballast (e.g. gym weights),
improving stability without encroaching on the boxer’s working zone.
The support structure holding the entire robot, particularly the
slewing bearing and linear stage, is
intentionally offset towards the rear
of the base (Figure 16b). This shifts the combined centre of
gravity backward, increasing resistance to forward tipping under
frontal punches.
Comparisons with alternative base shapes (Figure 16a) highlight that large
circular or rectangular bases tend to intrude into the boxer’s pivot space,
increasing the likelihood of the user stepping on or kicking the base during
movement. The trapezoidal layout, by contrast, maintains a compact yet
stable foundation, balancing usability, safety, and structural performance.
If subsequent weight estimates reveal that the base’s own mass is
insufficient to guarantee stability with an adequate safety factor, the rear
region of the base can be configured to accept
standard gym weights
as removable ballast. This approach keeps the design modular and adaptable
to different deployment environments.
Beyond executing realistic footwork drills, BoxBunny aims to remain
accessible to a broad demographic of users. To do so, the robot must allow
adjustment of the head and torso target heights to align
with different user statures. The height-adjustment system requirements are
summarised in the target metric specifications (Table 16) alongside key
selection criteria (Table 17).
5.3.3.2. Height Adjustment Mechanism Selection
Several concepts were generated to achieve the required stroke and load
capacity and were evaluated against the criteria above (Appendix B, Height
Adjustment section). The team evaluated three height-adjustment concepts. An
office-chair gas cylinder with guides was rejected because the gas lift
behaves like a spring–damper, causing vertical “bounce” and insufficient
stiffness under punches. An electric gas strut plus motorised guides allows
push-button adjustment but creates a complex, shared load path and is
unjustified for an adjustment that happens only occasionally.
The next concept consist of a manual screw jack with rear vertical linear
guides: the jack carries axial load and is self-locking, while the guides
take lateral and overturning forces. This concept offers high stiffness,
fail-safe behaviour, a simple load path, and acceptable manual effort,
making it the most suitable for BoxBunny. The main trade-off is that
adjustment is not as convenient or fast as a powered system. However, this
is acceptable given the low frequency of adjustment and the strong benefits
in stiffness, reliability, and simplicity. Consequently,
Manual Screw Jack with Rear Vertical Linear Guides
concept is selected as the height adjustment mechanism (Figure 17).
Having defined and selected the mechanisms for
footwork movement
, compact footprint, and
height adjustability
, the final step is to integrate these subsystems into a coherent lower
structure.
A key integration decision concerns the
relative ordering of the linear and rotational axes
. If rotation is placed above linear (Figure 18a), the
robot rotates about a point that is fixed along the linear axis. This
constrains the effective training area: rotation occurs around a single
longitudinal position, limiting variability in the spatial envelope of
targets. In contrast, if
linear motion is placed above rotation
(Figure 18b), the entire upper structure, including the height adjustment
and torso, can translate relative to the rotating base. This configuration
significantly
increases the reachable footprint of target positions
for the same base size, enhancing the richness of programmable drills.
Given the importance of varied footwork and angle creation, BoxBunny adopts
the second configuration, with linear motion above rotation
.
The final integrated stack, from bottom to top in Figure 19, starts from the
trapezoidal base providing a compact, stable footprint,
with space at the rear for optional ballast and structural anchoring. The
Slewing-ring yaw stage is mounted within the base,
comprising a four-point contact ball slewing ring with external gear, driven
by a pinion attached to a BLDC servo motor. The
Ball-screw-driven sliding table
on linear rails, mounted above the slewing ring, providing forward–backward
motion. The Manual screw jack with vertical linear guides,
mounted on the sliding table, providing height adjustment while resisting
lateral and overturning loads, and lastly, the
upper mechanisms
(torso, head, striking arms) attached to the output plate of the screw jack.
This integrated arrangement ensures that the
heaviest and most load-critical bearing
(the slewing ring) is located closest to the base, where it can best resist
overturning moments. The linear stage carries only the
loads associated with forward–backward motion and the mass above it, not the
full reaction of the base. The height-adjustment mechanism
is responsible primarily for axial support and is backed by rear linear
guides for lateral stiffness. Rotation and linear motion are fully
motorised
to allow programmable, repeatable training sequences, while height
adjustment remains a manual, self-locking operation for
safety and simplicity.
Overall, the integrated lower mechanism satisfies the key functional
requirements: it supports realistic, structured footwork drills; maintains a
compact and safe footprint; and offers adaptable target heights for a wide
range of users, forming a robust foundation for BoxBunny’s upper-body
interaction systems.
5.4. Robot Intelligence (Yogeeswaran)
The Robot Intelligence subsystem delivers Performance Analytics, Adaptive Fight Intelligence,
and Realistic Training Partner by using computer vision and an action-recognition model to detect punches,
predict responses, and analyse performance in real time. These product needs map to three core functions:
Reaction Time (system responsiveness), Human Tracking (accurate motion following), and Counter-punching (safe, reactive sparring without a partner).
5.4.1. Function Specifications
The subsystem must process vision data and trigger motor responses with latency comparable to a
human opponent. Based on reaction-time measurements from intermediate boxers, the system must meet
the following performance targets:
A depth camera can be used to meet these requirements as it provides reliable detection,
stable tracking, and accurate spatial awareness. Its wide field of view and high depth precision
enable consistent 3D user tracking, making it the most effective option for maintaining smooth and
dynamic real-time interaction between the user and the robot.
5.4.2. Computer Vision
5.4.2.1. Reaction Time and Human Tracking
Both functions use a pre-trained YOLO pose-estimation model, which provides continuous 3D keypoints
for real-time user tracking. The pose-estimation framework is based on the Ultralytics YOLO Pose model,
documented at
Ultralytics Pose Documentation.
Human Tracking: Relies on stable frame-to-frame detection for accurate motion prediction and safe interaction.
Reaction Time: Uses calibrated pose-change thresholds to detect punch initiation and trigger a response within the 150 ms target.
5.4.2.2. Counter Punch
The counter-punch function uses a custom-trained action-recognition model,
unlike reaction time and tracking, which rely on pre-trained YOLO pose estimation.
The model classifies punch types from pose sequences and maps them to predefined counteractions,
enabling dynamic and adaptive sparring behaviour. Here is the link to our
full training and inference codebase.
5.4.2.3. Model Pipeline
Training Pipeline: Pose estimation extracts keypoint sequences from labelled training videos.
These sequences are used to train a custom temporal model to recognise punch patterns. Various architectures (LSTM, TCN, transformer-based)
will be explored to balance accuracy and real-time latency.
Inference Pipeline: In real time, the depth camera captures frames, pose estimation generates keypoints,
and the trained model predicts the user’s punch. This predicted punch maps directly to a corresponding counteraction, allowing immediate robot response.
5.4.2.4. Trained Model
An initial LSTM-based model (Long Short-Term Memory,
a simple neural-network architecture designed to recognise patterns over time)
was trained on a small 2D keypoint dataset and achieved high accuracy on unseen videos, validating the end-to-end pipeline
from pose extraction to temporal classification. Next, a larger dataset will be collected using the robot’s depth camera to
incorporate 3D keypoints (x, y, depth), which is expected to significantly improve accuracy. The final model will also be
benchmarked for real-time inference speed to guide processor and hardware selection for meeting reaction-time constraints.
Jab Model
Hook Model
Uppercut Model
5.4.2.5. Position & Orientation of Camera (Testing)
Testing showed that classification accuracy decreases when the camera viewpoint differs from the
dataset’s original perspective, as the model fails to interpret unfamiliar pose patterns. This confirms
that the camera position on the robot must remain fixed, and future datasets must be captured from the mounted viewpoint to ensure consistent performance.
Top View
Bottom View
Front View
5.5. Control System (Yogeeswaran)
5.5.1.1. Component Overview
The diagram provides a simplified overview of the main connections within BoxBunny,
showing how power, data, and serial links are distributed across subsystems. Impact sensors on the
head and body mark key striking zones, while the left and right joints represent the robot’s active striking arms.
Parallel processing is essential for coordinating tasks such as vision inference, sensor feedback, and motor control in real time.
Common approaches for managing these parallel workloads include ROS 2, micro-ROS, and FreeRTOS. ROS 2 provides a modular, message-based
framework suitable for higher-level perception and decision making, while micro-ROS and FreeRTOS offer lightweight, deterministic task
scheduling on microcontrollers for low-latency motor and sensor loops. The system will adopt the appropriate approach depending on the
chosen processor and required timing precision, FreeRTOS for fast, time-critical control and ROS 2 for flexible integration of perception,
intelligence, and user-interface components.
5.5.1.2. Control System
A closed-loop control system is required to ensure BoxBunny can continuously correct its
motion based on real-time feedback, as shown in Figure 24. This allows the robot to respond
smoothly and safely to rapid load changes during training.
To achieve this, the motors can be driven using a PID-based controller such as the ODrive,
which implements a three-loop cascade:
the position loop outputs a target velocity,
the velocity loop outputs a target current, and
the current loop directly controls the motor torque.
This structure is necessary because each layer controls a different physical variable with different response speeds.
Torque reacts the fastest, so the current loop must run at the highest rate to stabilise the motor instantly.
Velocity changes more slowly, so the velocity loop runs on top to smooth motion.
Position changes slowest, so the outer loop updates least frequently to avoid oscillations.
By separating control across these time scales, the system remains stable, responsive, and safe under varying loads.
The PID loops handle critical edge cases encountered during sparring, including:
Sudden impacts (e.g., user punching the arm) – PID damps the disturbance and restores position smoothly.
User grabbing or holding the arm – torque limits prevent the motor from fighting the user.
User resisting or pushing against motion – current control avoids overcurrent and overheating.
Stall or near-stall conditions – prevents motor burnout by limiting torque.
By combining encoder feedback with this PID cascade, the robot maintains accurate movement
and robust protection during high-impact training.
6. Future Work (Zakir, Elgin, Jeanette, Yogeeswaran)
6.1 Test Plan
This section summarises the subsystem testing approach across mechanical, sensing, vision, control, and user interaction.
A full test plan will be finalised once the hardware is confirmed. Testing each stage separately ensures the system is validated
under realistic conditions and refined for safety, reliability, and responsiveness.
6.2 Project Plan
This section outlines the key milestones and overall timeline for
fabrication, subsystem testing, user testing, and full prototype assembly.
Buffer periods are included based on past project experience to accommodate
potential delays and ensure adequate time for troubleshooting and
refinement.
References
Boxing Mitts vs Boxing Sticks: What’s the Difference. (22 July, 2025). Retrieved from Bullsport Direct: https://www.bullsportsdirect.com/boxing-mitts-vs-sticks/
Cheraghi, M., Alinejad, H. A., Arshi, A. R., & Shirzad, E. (2014). Kinematics of Straight Right Punch in Boxing. Annals of Applied Sport Science, vol. 2, no. 2, 39-50.
Collective, W. (5 May, 2022). Why do people start training in Martial Arts? I asked a 1000 people to find out! Retrieved from Warrior Collective: https://warriorcollective.co.uk/blogs/wcarticles/why-do-people-start-training-in-martial-arts-i-asked-a-1000-people-to-find-out
G., D. T., Magdalena, I., Lindsay, W., I., D. D., & M., H. A. (2019). Understanding the Consequences of Repetitive Subconcussive Head Impacts in Sport: Brain Changes and Dampened Motor Control Are Seen After Boxing Practice. Frontiers in Human Neuroscience, Volume 13.
Illsin, A. (1 February, 2025). 10 Expert Tips for Choosing the Best Focus Mitts: A Complete Buyer's Guide. Retrieved from mmawarehouse: https://www.mmawarehouse.com/blogs/gear/best-focus-mitts
IxDF, I. D. (25 May, 2016). What are User Interviews? Retrieved from Interaction Design Foundation - IxDF: https://www.interaction-design.org/literature/topics/user-interviews
Lystad, P. R., Alevras, A., Rudy, I., Soligard, T., & Engebretsen, L. (2021). Injury incidence, severity and profile in Olympic combat sports: a comparative analysis of 7712 athlete exposures from three consecutive Olympic Games. British Journal of Sports Medicine, 1077-1083.
Stiller, J. W., Yu, S. S., Brenner, L. A., Langenberg, P., Scrofani, P., Pannella, P., . . . Postolache, T. T. (2014). Sparring and Neurological Function in Professional Boxers. Front Public Health, Volume 2.
Swenson, A. (4 January, 2022). Partner Training With Boxing Mitts, Pool Noodles, & Paddles. Retrieved from FightCamp: https://blog.joinfightcamp.com/boxing-equipment/partner-training-with-boxing-mitts-pool-noodles-paddles/
Appendix
1) Upper Mechanisms (Elgin)
a) Punch Execution
i) Joint Selection
To determine a tangible function specification for our system, we
established 2 key matrices through physical measurements and validations.
The fidelity of a real boxing match must be maintained. This dictates that
the mechanism must not obstruct critical striking zones (like the liver or
head) when in its neutral position. A mechanism positioned too low, for
example, would render uppercuts to the celiac plexus impossible, thereby
training the boxer to aim at an incorrect, clear target rather than the true
one.
Concept 1: Multi-Actuator Rig
The first concept involved the adoption of multiple actuators, one for each
desired strike. To cover both the left and right sides, this would require a
total of six motors. While this could theoretically provide striking
fidelity, the solution was deemed mechanically complex, and high cost.
Furthermore, the additional motors and mechanisms would create a bulky
profile, likely culminating in the obstruction of the user's striking
points.
Concept 2: Windmill Actuator
A second concept attempted to consolidate two strikes into one motor. It
utilized a "windmill" design, with a training stick extended on both ends of
a central motor. A downward rotation would execute a jab, and an upward
rotation would execute an uppercut. Although this design reduced the motor
count, its extended component for the uppercut would directly obstruct the
liver strike zone when in its default, neutral position. This constituted a
critical fidelity failure, as it would make one of the key strike zones
inaccessible.
Concept 3: 2-Degree-of-Freedom (2DOF) Actuator
The final design selection was a 2-Degrees-of-Freedom (2DOF) mechanism,
which was determined to offer the optimal balance of fidelity, complexity,
and cost by utilizing only two motors per arm. This 2DOF system is comprised
of two primary axes. The first is a pitch axis, which provides vertical
rotation and permits the execution of the standard downward jab. The
inclusion of a secondary yaw axis, providing horizontal rotation, enables
the system to execute both the horizontal hook and the diagonal uppercut.
Selection Criteria
Table A1 was created to aid the decision making for the appropriate concept
for the arm joint selection.
Concept 3 (2-DOF Actuator) is the only solution that meets the requirements
of high training fidelity and zero obstruction of the striking zones. The
significant challenge of its high control-system complexity is accepted as a
trade-off to achieve a functional and effective training product.
ii) Torque Calculations
Arm System Characteristics
Training Sticks
The training sticks, our target object of rotation, is firstly characterized
based on the existing polyethylene foam available to us. We first determine
the density of the current foam base on its dimensions and weight.
The specification for the new training sticks is based on the desired reach
of the robot.
Angular Speed and RPM
Our target matrix is to be able to match the striking tempo of a coach
during a training session. We have recreated a series of five 90° strikes
and recorded the total duration of these strikes.
Acceleration
The time it takes for the arms to accelerate to the desired speed is crucial
to determining the peak torque. A faster acceleration will result in a
larger peak torque. To accommodate the users, we base the acceleration time
on the human reaction time to visual stimulus, providing the users
sufficient time to react. The average human reaction time is typically 0.25
seconds to a visual stimulus (Crossley, 2021).
Arm System Torque Requirements
Static Torque
For rotation in a vertical plane, the maximum static torque happens as the
training sticks passes through the horizontal position (ф = 90). At this
point, the gravitational torque is at its maximum.
Acceleration Torque
To determine the acceleration torque required, we first model the training
sticks as a uniform rod to calculate its inertia.
The acceleration torque is the required to produce the angular acceleration
α.
Drag Torque
During the rotation of the training sticks, it also experiences aerodynamic
drag which contributes an opposite torque. To maintain a constant rotation,
the motor must apply a continuous torque to oppose the opposing drag. The
drag force is determined by the following equation:
Where C d is the drag coefficient, and ρ air is the
density of air. To calculate the resultant torque because of drag, we must
consider two other factors. Firstly, the velocity of the rotating pool
noodle is not a constant but varies with length from the pivoting point.
Secondly, the resultant torque must account for the increasing distance
across the pool noodle while still accounting for the increase in velocity.
To derive the appropriate equation, we first consider the infinitesimal
force along the pool noodle as a function of its length.
This force acts at a distance r from the pivot, so the infinitesimal torque
from drag is:
If we integrate from the pivot to the total length of the training sticks,
we can determine the equation to calculate the drag torque.
The coefficient of drag is also not a constant and is dependent on the
Reynolds number. We account for the highest Reynolds number which occurs at
the tip speed, assuming that there is no wind present.
Based on the calculations, the system has two distinct torque requirements.
Peak torque is the maximum, intermittent torque required by the system. It
occurs only during the 0.25-second acceleration phase while the arm is
simultaneously lifting against the maximum gravitational load (at the
horizontal position).
Continuous torque is the torque required to maintain the arm's operation
after acceleration is complete. This load is highest when the arm is moving
through the horizontal position at full speed, where the motor must fight
both gravity and aerodynamic drag.
Final Arm System Specifications
It is standard engineering practice to apply a safety margin to account for
modelling inaccuracies, unforeseen friction, variations in material
properties, and voltage fluctuations. A safety factor of 1.5 is applied to
the required torque.
iii) Motor Selection Criteria
Table A8 was created to aid the decision making for the appropriate motor
for the arm joints.
2) Lower Mechanisms (Jeanette)
a) Footwork Movement
i) Motion Axis Selection
Having to accomplish two motions, decision has to be made to have both
controlled within one mechanism or decouple those axes into separate
control. One possible collective system is using omnidirectional wheels. The
omnidirectional wheel concept provides full planar mobility of the robot
base. A single drive system enables translations in both the X and Y
directions, as well as on-the-spot rotation, by appropriately commanding the
individual wheel speeds. At the control level, this appears attractive
because any desired planar velocity (forward, lateral, yaw) can be
synthesised from the same set of actuators.
Benefits of Omni-wheels for Boxing Application
These advantages are well aligned with general mobile robotics applications,
where the primary design objective is flexible navigation in cluttered
environments.
Limitations of Omni-wheels for Boxing Application
However, this comes with limitations for a punch-receiving boxing robot. For
a boxing trainer robot that must repeatedly absorb impulsive loads from
punches while maintaining a stable stance, several limitations arise when
using omnidirectional wheels:
Most boxing gyms do not provide the ideal hard, polished surface assumed in
many mobile robotics applications. Instead, they commonly use interlocking
foam mats, rubber tiles, or sprung or cushioned flooring systems.
On these surfaces:
Wheels can sink slightly into the surface, changing the effective
rolling radius and increasing rolling resistance.
The floor itself adds another compliant layer beneath the robot,
amplifying pose creep under repeated impacts.
Small shifts at the base accumulate over rounds, making it
difficult to maintain a consistent stance reference relative to the
user.
Consequently, maintaining repeatable stance alignment, which is critical for
structured drills and quantitative training metrics, is challenging with an
omnidirectional base on typical gym flooring.
Benefits of Decoupled Axes for Boxing Application
A design that explicitly decouples the primary motions, e.g. a linear rail
for the forward–backward axis combined with a slewing bearing for rotation,
sacrifices omnidirectional mobility but introduces significant advantages
for this application:
Selection Matrix – Forward-Backward Range Motion
Table shows the decision matrix created to aid the linear footwork movement
axis selection, assuming rating scale 1-5 (5 = best performance for
BoxBunny’s needs). Omnidirectional wheels score better on cost/simplicity as
one system that does everything. The linear rail is much stronger on
stiffness, repeatability, and independence from floor compliance, which are
critical for a punch-receiving training robot. Thus, for forward-backward
motion, the decoupled linear rail is the preferred solution.
Selection Matrix – Yaw Rotation (On-Spot Turning)
Table B10 shows the decision matrix created to aid the rotational footwork
movement axis selection. Omnidirectional wheels can rotate on the spot, but
yaw stiffness and angle repeatability are limited by wheel slip and floor
compliance. A slewing bearing driven by a motorised pinion offers high
rotational stiffness, well-defined backlash, and high moment capacity, all
independent of floor conditions. For on-spot rotation, the decoupled
rotating base is the preferred solution.
Motion Axis Selection for BoxBunny Footwork Movement
Across both motions, the selection matrices show that omnidirectional wheels
are attractive for general mobile robots but are penalised here by low
stiffness, pose creep, and strong dependence on floor conditions. Decoupled
axes (linear rail + slewing bearing) provide the stiff, repeatable,
floor-independent geometry needed for a boxing trainer that must reliably
hold stance under repeated punches. Therefore, for this application, the
project adopts a decoupled-axis architecture: Motorised linear rail for
forward–backward range, and slewing bearing with external gear and
pinion-driven motor for yaw rotation.
ii) Linear Motion Selection
(1) Power Transmission Drive Selection
The power transmission drive for the linear stage must reliably move the
entire BoxBunny robot (≈ 20–25 kg including arms and torso) forward and
backward over a short stroke (~300 mm), while withstanding dynamic punching
loads. Key requirements for this axis are:
Load and thrust capacity: Sufficient to
accelerate the robot mass and resist horizontal reaction forces
during punches.
Positional stiffness and repeatability: The
base must return to the same stance positions for drills without
noticeable drift.
Moderate travel speed: Fast enough to
simulate a boxer stepping in/out, but not a high-speed shuttle
(short strokes, <1 m).
Robustness under repeated duty: Frequent
starts/stops and impulse loads during training sessions.
Compact integration: Drive must fit under the
rotating base and vertical structure without excessive footprint.
Three common drive types were evaluated: lead screw,
ball screw, and timing belt.
Lead Screw Drive
Lead screws convert rotary motion to linear motion via a threaded rod and
nut. They are generally low cost, mechanically simple, and
can use self-locking polymer or anti-backlash nuts, which reduces drift in
vertical axes and improves repeatability for light loads. However, for
BoxBunny’s application they present several drawbacks:
Lower efficiency and higher friction, leading
to more heating at higher speeds or duty cycles.
Limited load and speed capability with typical
polymer nuts; not ideal when combining moderate loads, repeated
motion and dynamic impacts.
Moderate accuracy and life compared to ball
screws, especially under higher thrust and duty.
Lead screws are therefore more suited to light-load, low-speed or low-duty
applications, and were not selected as the primary drive for BoxBunny’s main
footwork axis.
Ball Screw Drive
Ball screws use recirculating balls between the screw and nut, providing
rolling contact instead of sliding. For BoxBunny, they
offer several key advantages:
High thrust and load capacity appropriate for
moving a ~20–25 kg robot and resisting horizontal reaction forces.
High stiffness, accuracy and repeatability, due to
low backlash and well-controlled preload, which supports repeatable
stance positions.
High mechanical efficiency, reducing motor
torque requirements and heat build-up under frequent moves.
Good durability under repeated duty, making
them suitable for regular training sessions with many cycles.
The main limitations are higher cost and the need for
periodic lubrication, but these are acceptable trade-offs
given the short stroke and performance requirements.
Belt Drive
Belt-driven actuators use a toothed belt and pulleys to generate linear
motion. Their strengths lie in:
High speed and long travel capability,
typically over several metres.
High efficiency and relatively simple maintenance
for long-stroke, lower-load systems.
However, for BoxBunny’s short-stroke footwork axis they are less suitable:
Lower positional stiffness and accuracy compared
to screw drives, due to belt elasticity.
Susceptibility to elongation and shock loads,
which can degrade repeatability under punching impacts.
Need for periodic belt re-tensioning, and
potential back-driving / drift in horizontal or
inclined applications.
Given the short stroke and high stiffness requirement, these disadvantages
outweigh the speed benefits.
Selection Matrix – Power Transmission Drive in Linear Rail
Table shows the decision matrix created to aid the power transmission drive
in linear rail selection.
Drive Selection for BoxBunny Linear Stage
Considering BoxBunny’s specific requirements of moderate stroke, significant
mass, repeated dynamic loads, and the need for high positional stiffness and
repeatability, the ball screw drive is selected for the
linear stage. It provides:
Sufficient thrust and load capacity for the full robot mass,
High stiffness and repeatability for consistent stance positions,
and
Robust performance under repeated use, with manageable lubrication
requirements.
This makes a ball screw–driven linear rail the most
appropriate power transmission solution for the BoxBunny footwork axis.
(2) Motor Selection
The motor driving the ball-screw linear stage must translate the full
BoxBunny mass (~20–25 kg) over a short stroke (≈200–300 mm) while
maintaining precise, repeatable stance positions under punching
disturbances. Key requirements are:
High torque at low–medium speed to drive the ball
screw without stalling.
High positioning accuracy and repeatability so
stored “stance distances” are consistent.
Good disturbance rejection and stiffness to resist
position loss when the robot is punched.
Realistic cost and integration complexity for
a prototype.
Three motor families were evaluated:
Open-Loop Stepper Motor (NEMA-Frame)
, Closed-Loop Stepper / Hybrid Servo (Stepper + Encoder),
and Industrial AC/BLDC Servo Motor (Servo Pack).
Open-Loop Stepper Motor
Standard NEMA-frame stepper motors, driven in microstepping mode, are widely
used in low-cost linear stages.
Advantages
High holding torque at low speed, matching
the ball screw’s operating range.
Simple step/direction interface; no encoder
required.
Very low cost and excellent availability of
motors, drivers, and mounting hardware.
Limitations for BoxBunny
Open-loop control: missed steps under sudden loads
or aggressive acceleration are not detected.
Loss of steps leads to silent position drift,
which directly undermines stance repeatability.
Resonance and vibration can occur if not carefully
managed.
Open-Loop Stepper Motor is therefore an attractive baseline for cost and
simplicity, but its vulnerability to undetected position loss is a concern
for a punch-receiving system.
Closed-loop steppers (often sold as “hybrid servos”) combine a stepper motor
with an integrated encoder and a driver that closes the position/velocity
loop.
Advantages
Encoder feedback allows detection and correction of missed steps,
improving reliability under shock loads.
Similar NEMA form factor and mounting as a standard stepper;
easy mechanical integration with the ball screw.
Retains high low-speed torque while significantly
improving positioning accuracy and stiffness.
Many drives still accept simple step/direction commands,
minimising software changes
.
Limitations for BoxBunny
Higher cost than open-loop steppers.
Some tuning of current limits and loop gains may
be required, though typically
simpler than full industrial servos
.
Closed-Loop Stepper / Hybrid Servo strikes a good balance between
performance and complexity for BoxBunny’s linear stage.
Industrial AC/BLDC Servo Motor (Servo Pack)
Industrial servo systems pair an AC/BLDC motor with a high-resolution
encoder and a matched servo drive.
Advantages
Excellent torque–speed characteristics and disturbance rejection.
Very high positioning accuracy, repeatability, and configurable
stiffness.
Well suited to demanding industrial positioning applications.
Limitations for BoxBunny
Highest cost among the three options (motor, drive, cables).
Greater integration complexity: power wiring, controller interface,
and tuning of control loops.
Over-specified for a short-stroke, low-speed axis in a student
prototype unless hardware is already available.
Industrial AC/BLDC Servo Motor offers the best raw performance, but at a
cost and complexity level that may not be justified for this application.
Selection Matrix – Motor in Linear Rail
Table shows the decision matrix created to aid the motor in linear rail
selection.
Motor Selection for BoxBunny Linear Stage
Considering BoxBunny’s need for
high stiffness and repeatability under punching disturbances
, while keeping integration feasible for a student project,
Closed-Loop Stepper / Hybrid Servo
is selected to drive the ball-screw linear stage. It provides a substantial
improvement in robustness over an open-loop stepper without the full cost
and integration burden of an industrial servo pack.
iii) Rotational Motion Selection
(1) Bearing Selection
The yaw rotation system must support the full BoxBunny robot mass (≈ 20–25
kg), plus dynamic punching loads and overturning moments from the elevated
torso and arms. The bearing must:
Carry combined axial, radial and moment loads from
punches and body weight.
Provide high rotational stiffness so the robot does not
“wobble” or twist when struck.
Remain durable under repeated shock and vibration.
Allow reasonably precise positioning when driven by an
external pinion and motor.
Be cost-effective and readily available for a student
prototype
· Fit within limited height of the lower structure
and interface cleanly with the rotating base plate.
Four bearing families were evaluated: a low-cost
lazy-Susan bearing
and three types of slewing-ring bearings.
Lazy-Susan Bearing
Lazy-Susan (turntable) bearings are inexpensive, thin-section assemblies
commonly used in furniture and light-duty turntables.
Advantages
Very low cost and widely available.
Simple mounting between two plates.
Compact axial height.
Limitations for BoxBunny
Intended mainly for light axial loads, with
limited overturning-moment capacity.
Minimal preload → noticeable play in tilt and
rotation; unsuitable for precise yaw positioning.
Not designed for repeated impacts; looseness and
wear are expected under punching loads.
Lazy-Susan bearing is mechanically attractive for a toy turntable, but does
not offer the stiffness, moment capacity, or durability required for
BoxBunny.
Slewing-ring Bearing
Four-Point Contact Ball Slewing Ring Bearing
Four-point contact ball slewing rings use a single row of balls in a
specially profiled raceway, allowing each ball to carry load at up to four
contact points.
Advantages
Designed to support
combined axial, radial and overturning moment loads
, including varying and shock loads.
Can be preloaded to reduce internal clearance and
improve stiffness.
Lower friction than cross-roller designs → adequate speed with
smooth motion (although high speed is not critical here).
Widely used in light–medium duty slewing applications,
readily available in standard sizes
and generally cheaper than cross-roller slewing
rings.
Provides sufficient stiffness and load capacity
for a 20–25 kg robot with moderate punching loads, with safety
margin.
Limitations for BoxBunny
Rotational stiffness is lower than cross-roller
bearings for the same size and preload; small elastic deflections
are still present under heavy impact.
Thisoffers a good balance of load
capacity, stiffness, availability and cost, and satisfies BoxBunny’s yaw
requirements with adequate margin.
Cross-Roller Slewing Ring Bearing
Cross-roller slewing rings use cylindrical rollers arranged in alternating
orientations (X-pattern) between inner and outer rings.
Advantages
Very high tilting and rotational stiffness due to
line contact and crossed layout.
Excellent for
combined axial, radial and moment loads
in compact envelopes.
Commonly used in
robot joints and precision rotary stages
requiring minimal play and high rigidity.
Limitations for BoxBunny
Generally, more expensive than four-point ball slewing
rings.
Somewhat less readily available in suitable sizes from
mainstream suppliers compared to four-point ball types.
Higher performance than strictly required for this application.
This is technically the stiffest option, but its
cost and availability
are not justified when Four-point contact ball slewing rings already meets
the design requirements.
Thrust Ball Slewing Rings
Thrust ball slewing rings are optimised mainly for
axial loads
, with limited radial and moment capacity.
Advantages
Efficient support for vertical loads with low
friction.
Limitations for BoxBunny
Not optimised for large overturning moments from
an elevated torso and punching impacts.
Usually requires additional bearings to handle
radial and moment loads, increasing complexity.
This does not match the combined-load and high-moment demands of BoxBunny’s
yaw joint without extra components.
Selection Matrix – Bearing in Yaw Rotation System
Table shows the decision matrix created to aid the bearing in yaw rotation
system selection.
Bearing Selection for BoxBunny Yaw Rotation
Given that Four-Point Contact Ball Slewing Ring
Bearingalready meets BoxBunny’s load and stiffness requirements
with margin, and offers better
cost-performance and procurement practicality
, it is selected as the yaw bearing for the system.
(2) Gear Reduction Selection (Baseline)
The gear reduction for the yaw axis is largely dictated by the choice of
slewing ring bearing with external gear teeth. Once the
bearing diameter and tooth form are fixed, the mating pinion and resulting
gear ratio follow naturally. The calculations in this section are presented
as a baseline template: when the final torque, speed, and load analysis for
BoxBunny are complete, the same steps can be repeated with updated values.
Example model: LILY Bearing MTE-265X (externally geared)
Gear mesh: Module 5 external gear
From the bearing datasheet, the external ring (driven gear) is specified
with:
Driven gear (slewing ring): 84 teeth
Driving gear (pinion): 17 teeth
The pinion is mounted on the motor shaft (or via a short shaft + coupling),
driving the external gear on the slewing ring.
Gear Ratio Calculation
The gear ratio (GR) is defined as:
where
Ndriven = number of teeth on the external slewing ring,
Ndriving = number of teeth on the pinion.
For this configuration:
Ndriven = 84
Ndriving = 17
This means the pinion (and motor) must rotate approximately
4.94 times faster than the slewing ring.
Baseline Motor Speed Requirement
Assume a target yaw speed at the slewing ring of:
Desired output speed (slewing ring):
The required motor (pinion) speed is:
Convert to RPM:
So, for an output yaw speed of 150°/s, the motor driving
the 17-tooth pinion must run at approximately 124 rpm with
this gear ratio. This is a
reasonable speed range
for many commercial motors and servo systems. When final requirements are
known, the same equations can be reused with updated ωout and tooth counts.
Baseline Torque Relationship
Once the required yaw torque at the slewing ring is known from the load
analysis (see below), the required motor torque can be
estimated as:
where
Tslewing = required torque at the slewing ring (from loads),
GR = 4.94 (gear ratio),
η = efficiency of gear mesh (typically 0.90-0.95 for a well-lubricated spur gear pair).
This formula provides the baseline motor torque requirement once the yaw
moment demand has been quantified.
Load Calculation
To size the slewing ring and motor torque properly, the main loads on the
yaw bearing must be identified. Figure B1 illustrates the free-body diagram
of loads acting on the bearing.
Fa: Axial Load
Total vertical load acting along the axis of rotation.
Primarily the
weight of all rotating components
above the bearing (torso frame, arms, padding, sensors, etc.).
Fr: Radial Load
Forces acting
perpendicular to the axis of rotation
in the horizontal plane.
Includes horizontal components of punching forces
and any off-centre loads due to asymmetric mass distribution.
Mk: Tilting Moment (Overturning Moment)
Moment about a horizontal axis through the bearing centre.
Arises from:
The weight of the robot’s upper structure
acting at a vertical offset from the bearing centre
(lever arm).
Punching forces applied away from the bearing centre
(e.g. at the head or torso pad).
Typically, the most critical load for a
slewing ring.
Calculated from the centre of gravity (CG)
location and impact points:
where F is the force (weight or impact) and d is the perpendicular distance from the bearing centre to the line of action.
Impact/Shock Loads
Boxing involves dynamic impacts. A peak punch
force (e.g. up to ~1.5 kN for an average user) acting at a
certain radius from the bearing centre will generate a short
duration tilting moment.
Such impacts are typically accounted for using a
service factor (SF)
, multiplying the nominal static/dynamic loads by SF to ensure
the bearing and gear are sized with sufficient margin.
Mr: Friction Torque
Internal resistance to rotation within the bearing assembly,
dependent on bearing diameter, preload, rolling element type,
and lubricant.
Contributes to the baseline torque the motor
must overcome even without external loads.
These load components form the input to the manufacturer’s
rating charts (axial load, radial load, tilting moment, friction torque).
Once the final BoxBunny mass distribution and impact assumptions are fixed,
the above framework can be used to:
Compute Fa, Fr, and Mk
Check the slewing ring against its
static and dynamic load ratings
.
Derive the required yaw torque at the bearing, and
from that, the motor torque using the gear ratio
and efficiency relations described earlier.
(3) Motor Selection
The yaw motor, together with the pinion and slewing ring, must rotate the
full BoxBunny upper body about the vertical axis while resisting
punching-induced moments. Key requirements are:
Sufficient torque at the pinion to overcome:
Tilting moment transmitted through the gear mesh
Bearing friction torque
Inertial torque during acceleration/deceleration
Target speed: ≈150°/s at the slewing ring (≈124
rpm at the motor with GR ≈ 4.94).
Accurate, repeatable positioning for defined yaw angles
(e.g. open/closed stance modes).
Good disturbance rejection so the robot does not
“twist” or lose position under impacts.
Reasonable cost and manageable integration for a
student prototype.
Three realistic motor options were evaluated:
Brushed DC Gearmotor
, Closed-Loop Stepper / Hybrid Servo, and
BLDC Servo Motor with Encoder
.
Brushed DC Gearmotor
Brushed DC motors with an integrated gearhead are common in low-cost rotary
stages.
Advantages
Simple control: speed is roughly proportional to
applied voltage; direction via polarity.
High torque at low output speed due to the gearhead,
potentially matching the required 124 rpm.
Widely available and relatively low cost.
Limitations for BoxBunny
Poor inherent position control: without an encoder and
closed-loop controller, position is estimated indirectly (via
time/voltage), which is not precise enough for repeatable yaw angles.
Even with an encoder added, most low-cost gearmotors exhibit
gearbox backlash
, causing play at the pinion and visible yaw “looseness” when
punched.
Brushed motors have brush wear and may require more
maintenance in long-term use.
Thisis attractive for torque and simplicity, but the
combination of backlash and weaker position control makes it unsuitable as
the primary yaw actuator for BoxBunny.
Closed-Loop Stepper / Hybrid Servo
A closed-loop stepper (hybrid servo) combines a stepper motor, encoder and
smart driver. It accepts step commands but internally closes the loop on
position/velocity.
Advantages
Good low-speed torque, suitable for directly driving
the pinion through the 4.94:1 reduction.
Encoder feedback allows detection and correction of position error
under disturbances.
Relatively simple command interface (often still step)
compatible with typical motion controllers.
More compact and affordable than many industrial AC servos.
Limitations for BoxBunny
Steppers have torque ripple and microstep non-linearity
, which can produce small angle “cogging” in very slow or static
positions; this can be felt as slight chatter when the user
manipulates the robot.
Dynamic performance is good but not as smooth as a
sinusoidally-commutated BLDC servo under rapid direction changes.
Acoustic noise and vibration can be higher than BLDC at certain
operating points.
This is a solid option with closed-loop control and adequate torque, but
its torque ripple and dynamic smoothness are not ideal for a “human-facing”
yaw motion that should feel smooth and responsive.
BLDC Servo Motor with Encoder
This option uses a brushless DC motor paired with a
rotary encoder and a FOC/servo driver that
closes the loop on current, velocity and position (i.e. a BLDC servo).
Advantages
Smooth, sinusoidal torque production via field-oriented
control, reducing cogging and vibration at low speed.
Good torque–speed envelope easily meets ≈124 rpm at the
motor with margin for faster moves if desired.
Closed-loop position control using encoder feedback →
high yaw accuracy and repeatability, with active disturbance
rejection under punching.
High efficiency and low maintenance, since there are no
brushes to wear out.
Well suited to continuous rotary motion with frequent accelerations
and reversals.
Limitations
Higher integration complexity than a simple DC
gearmotor: requires encoder wiring and a suitable BLDC servo
driver.
Typically, more expensive than basic brushed motors,
though often comparable to or slightly above closed-loop steppers
depending on brand.
This offers the best combination of
smoothness, torque, precision and disturbance rejection
for BoxBunny’s yaw axis, at a complexity level that is still manageable for
a student build using off-the-shelf BLDC servo drivers.
Selection Matrix – Motor in Yaw Rotation System
Table shows the decision matrix created to aid the motor in yaw rotation
system selection.
Motor Selection for BoxBunny Yaw Rotation
For the yaw axis, user perception and stability are critical: the robot
should rotate smoothly and feel “solid” under punches, without rattling or
drifting from its commanded angle. While brushed DC gearmotors and
closed-loop steppers can meet basic torque and speed needs, their drawbacks
in backlash and torque ripple/noise make them less suitable.
Given these requirements, BLDC Servo Motor with Encoder is
chosen as the yaw motor. It provides smooth and controlled yaw motion, high
positional stiffness and repeatability, and robust behaviour under
impact-induced disturbances, all of which align well with the BoxBunny
design goals.
b) Height Adjustment
The height adjustment system must vary the BoxBunny torso/head height by
approximately 300–400 mm to accommodate different user
heights, while:
Carrying the axial load of the upper structure
(~20–25 kg plus padding and arms)
Resisting lateral forces and overturning moments from
punches
Remaining stiff and “non-bouncy” during training (once
set, the height should not oscillate)
Being safe and simple for users to adjust occasionally
(not continuously during a round)
Fitting within the available vertical envelope above the linear
stage
Three concepts were explored.
Concept #1: Office Chair Mechanism + Vertical Linear Guide
The first idea was to adapt the office chair mechanism
(Figure B2): a gas lift cylinder to carry the axial load,
combined with vertical linear guides at the back to resist
lateral punching forces.
Advantages
Gas lifts are readily available and compact.
Can carry the static axial load of the robot mass.
Vertical linear guides at the back can be sized to take
lateral loads and overturning moments
.
Limitations for BoxBunny
Gas cylinders are inherently compliant – they
behave like a spring/damper. Under repeated punches, this leads to
vertical “bouncing” of the torso, which feels
unrealistic and reduces strike rigidity.
The internal gas spring force varies with stroke; stiffness is
not easily tuned for the specific mass and dynamic
loads.
Conventional office-chair gas lifts are designed for
human seating comfort
, not for resisting repeated impact loads at an elevated lever arm.
Conclusion: While mechanically simple, Concept #1 does not
provide sufficient vertical stiffness; the gas lift
introduces unwanted bounce under impact.
Concept #2: Electric Gas Strut + Motorised Vertical Linear Guide
The second concept combined an
electrically actuated mechanism
for height control with a gas strut at the bottom to assist
in carrying axial load (Figure B3):
A motorised drive (e.g. screw or belt) to move the
upper carriage along vertical linear guides.
A gas strut mounted beneath to partially support
the robot’s weight.
Advantages
Allows push-button height adjustment, potentially even
mid-session.
The gas strut can reduce the required motor torque by
counterbalancing
part of the robot’s weight.
Vertical linear guides again provide
lateral and moment support
.
Limitations for BoxBunny
Mechanically and electrically complex:
coordinating motor drive, gas strut support, and guide loads is
non-trivial.
Multiple elements share the load path (motor, screw/drive, gas
strut, guides), making
stiffness and reliability harder to predict
.
For an application where height is adjusted
occasionally
(between users, not continuously), the added complexity and cost of
powered adjustment is not justified.
Conclusion: Concept #2 offers convenience but at the cost
of high complexity and a more difficult load path. It is
unnecessarily sophisticated for a height setting that is changed only
occasionally.
Concept #3: Manual Screw Jack + Vertical Linear Guides
The final concept uses a manual screw jack with handwheel
mounted on top of the linear stage, combined with
vertical linear rails
as the primary guides.
The screw jack carries the axial load
of the robot and provides precise height adjustment via handwheel
rotation.
Two vertical linear rails (each with multiple long
blocks) mounted behind the robot carry
lateral forces and overturning moments
from punches, ensuring the structure does not tilt.
Advantages
High axial load capacity and stiffness: screw
jacks are designed for heavy-duty industrial applications; they
carry the full weight with minimal vertical compliance.
Self-locking: with an appropriately chosen screw
lead and efficiency, the jack is non-backdrivable, so
the height does not drift or bounce under impacts. No extra brake
is required.
Simple load path: axial loads go directly
through the screw jack; lateral and moment loads are taken by the
vertical rails → easier to analyse and design.
Manual but ergonomic: height is adjusted via
a handwheel; this is acceptable because height changes occur
between users or rounds, not continuously.
Mechanically robust and maintainable, using
standard industrial components (jack + linear guides).
Limitations
Requires manual effort to adjust height; not as
convenient as push-button systems.
Adjustment speed is slower than an electric actuator, though still
acceptable for occasional use.
Conclusion:
Concept #3 provides high holding stiffness, self-locking behaviour
, and a clear separation of axial vs lateral/moment support
. It meets BoxBunny’s functional requirements with moderate cost and
manageable complexity and is therefore selected as the height adjustment
solution.