Back | Home

Standards

WAISF Guideline for Safety, Technical Methods and Use of Embodied-AI Sports Robots

2026-10-27

World AI Sports Federation (WAISF)
Guideline for Safety, Technical Methods and Use of Embodied-AI Sports Robots
Applicable to Taekwondo, Boxing, Kendo, Lacrosse and Cricket Systems

Document Title: WAISF Guideline for Safety, Technical Methods and Use of Embodied-AI Sports Robots
Document No.: WAISF-GDL-2026-001
Issuing Body: World AI Sports Federation (WAISF)
Version: 1.0
Publication Date: 27 October 2026
Status: International industry guidance. Unless expressly adopted by applicable law, a competent authority, a sport governing body or a contract, this document is not itself a national law, mandatory technical regulation or accredited international standard.
Recommended Review Cycle: Every 12 months, and earlier following a serious incident, material technological change or change in applicable law.

1. Purpose and Core Principles

This document provides guidance for national sport governing bodies, clubs, event organizers, robot manufacturers, system integrators, AI developers, laboratories and research institutions on the safe development, testing, training, demonstration and competition use of embodied-AI sports robots. It addresses risks arising from movement, impact, crushing, loss of control, perception errors, data processing and cybersecurity.

The guidance may apply to humanoid platforms, including Unitree, EngineAI and other platforms, as well as purpose-built ball-launching equipment. No brand or model is automatically approved for safe use because of its name, advertised specifications or previous demonstrations.
1.1 Human safety first
Human life, bodily integrity, informed consent and the right to stop or withdraw take priority over competitive performance, system capability, entertainment, commercial interests and model-training objectives. “Research,” “demonstration,” “voluntary participation” or the claim that a robot does not fatigue does not remove the duty to control risk.
1.2 Risk-based classification and authorization
A robot must be assessed according to its actual task, output capability, contact method, operating environment and degree of human exposure. Risk classification must not be based solely on a robot model or software-mode label.
1.3 Independent safety layer
Safety functions must not depend entirely on the same AI model, vision channel or network connection. Where practicable, critical force limits, speed limits, emergency stops, isolation and energy-control functions should be implemented in an independent and verifiable control layer.
1.4 Human supervision and stoppability
All physical human-robot training and demonstrations must be supervised by a trained on-site lead. Any athlete, coach, referee or safety lead may request an immediate stop. The system may be reset only after the hazard has been assessed and controlled.
1.5 Evidence over assurances
Simulation success, a successful demonstration, the presence of an emergency stop or soft external padding does not by itself establish safety. Safety conclusions should be supported by traceable risk assessments, measured data, test records, fault-injection results and independent review.

2. Scope and Activity Categories

This document covers:

1. Taekwondo: Robot-held targets, athletes kicking robot target pads, robot-provided constrained movement cues or defensive responses, approved controlled human-robot technical training, and robot-to-robot demonstrations or contests.
2. Boxing: Robot-held mitts, athletes striking robot-held targets, constrained moving targets or defensive cues, human-robot training, and robot-to-robot training or contests.
3. Kendo and other safety-adapted training with traditional weapons: Action recognition, distance training, timing drills and robot-to-robot demonstrations using approved training shinai, foam equipment or other risk-assessed blunt training implements. This document does not authorize real swords, sharp implements or uncontrolled high-speed rigid weapons near people.
4. Lacrosse: Robot ball launching, passing, catching, mobile passing-and-catching, goalkeeper and target drills, including ball interactions between humans and robots or between robots.
5. Cricket: Automated or robotic bowling machines, control of ball speed and trajectory, batting practice, catching practice and facility safety management.

The following operating categories are distinguished:

- Class A — Simulation only: Operation occurs only in a digital simulation, with no physical exposure of people.
- Class B — Enclosed robot-to-robot (R2R): Physical robots operate against each other while people remain outside an exclusion zone, with protective boundaries, remote emergency stops and protection against debris or balls.
- Class C — Non-contact human-robot training: People and robots share a training environment, but bodily contact is not intended. Examples include motion demonstration, tracking, fixed-path ball launching and controlled moving targets.
- Class D — Controlled-contact human-robot training: Contact may occur with a person, wearable equipment or training apparatus. A task-specific risk assessment, validation and authorization are required.
- Class E — Human-robot competition or high-intensity contest: A robot operates autonomously or semi-autonomously in a task involving speed, force, reaction or adversarial interaction with a person. This is the highest-risk category. Physical contact competition must not proceed without a dedicated independent assessment, validation and written authorization.
- Class F — High-speed ball launching: Lacrosse or cricket launchers and similar projectile equipment require a dedicated launch zone, personnel separation and mechanical/electrical safety controls.

R2R does not mean risk-free. Robot-to-robot operation is permitted only when risks from personnel exposure, collision, flying parts, falls, fire, batteries and remote loss of control are adequately controlled.

3. Risk Assessment and Activity Authorization
3.1 Pre-operation risk assessment
A documented risk assessment is required for each venue, robot model, end effector, control-software version, participant group and operating mode. At minimum, assess:

- Collision, striking, kicking, crushing, shearing, entanglement, falls and compression;
- Force, speed, kinetic energy, stiffness, reach and stopping distance of joints and end effectors;
- Vision occlusion, sensor failure, latency jitter, communication loss, localization drift and synchronization errors;
- AI misclassification, abnormal policies, reward-induced hazardous behavior, model updates and human/robot identity errors;
- Emergency-stop failure, software crashes, controller restarts, power faults, battery thermal events and mechanical damage;
- Exposure of athletes, coaches, spectators, minors, people with disabilities and bystanders;
- Privacy, cybersecurity, personal data, video and biometric-data processing.
3.2 Hierarchy of risk controls
Prioritize elimination of hazards, personnel separation, physical guarding, independent hardware limits, validated safety controls, operating procedures and personal protective equipment. Warning signs and training do not replace feasible engineering controls.
3.3 Restrictions where assessment is incomplete
A system that has not completed risk assessment, acceptance testing or emergency drills may only undergo necessary engineering tests under controlled low-risk conditions. It must not be deployed directly in public training, physical human-robot contact or competition.

4. Robot Hardware and Mechanical Safety
4.1 Limits on force, speed and range of motion
- Set task-appropriate limits for speed, acceleration, joint torque, current, end-effector force/torque and range of motion.
- In human-robot mode, safety limits must be determined through risk assessment and measurement, with applicable standards considered, and implemented at the low-level control layer. An upper-level AI policy’s promise to comply is insufficient.
- A single impact-force value must not be treated as a universal safety threshold for every body region, contact area, pulse duration, robot mass or speed.
- Consider end-effector mass, contact geometry, velocity, stiffness, collision duration, rebound, secondary impacts and the participant’s protective equipment.
- Mechanical limits, software limits and safety controls should provide multiple layers of protection where practicable. Software limits must not be the sole safeguard.
4.2 Soft covering and crush-point protection
- Parts that may contact a person should use tested energy-absorbing coverings, rounded edges and anti-crush designs.
- Padding thickness alone does not establish safety. Validate energy absorption, durability, detachment risk, cleaning and maintenance, and effects on joint movement.
- Guard or otherwise engineer out crushing and shearing points at joints, linkages, belts, gears, grippers, elbows, knees and between the robot and walls or equipment.
- If a guard is damaged, loose or interferes with a safety sensor, the equipment must be taken out of service until repaired and re-inspected.
4.3 End effectors and training equipment
- Boxing mitts, Taekwondo body targets, training swords, lacrosse heads and cricket-launching equipment must be assessed for secure attachment, durability and failure modes.
- Human-contact training must not use sharp, rigid, brittle or penetration-capable end effectors unless a task-specific risk assessment establishes an appropriately controlled and lawful use. Unassessed equipment is prohibited.
- Kendo activities must use clearly identified training implements approved by the project safety lead. Separate rules must address target zones, speed, implement stiffness and protective equipment. Real swords must not be used for human-contact training.
- Helmets, chest protectors, gloves, eye protection and other sports protective equipment must comply with applicable sport rules. Protective equipment does not authorize increased robot output.
4.4 Emergency stop and safe energy control
- Every physical system must have clearly identifiable emergency-stop devices that can be reached quickly. The on-site lead must be able to trigger an emergency stop without entering the hazard zone.
- High-risk systems should have an emergency-stop path independent of the main AI and ordinary communications link. Wireless stops may supplement but must not be the sole stop mechanism unless that design has been validated for the intended risk.
- Define the safe state after an emergency stop: halt hazardous motion, control residual energy, manage limbs or arms that could fall, and prevent unexpected restart.
- Cutting power is not always the safest response by itself. Systems that could fall, become unstable or release a held object when de-energized require a validated safe-stop and energy-control design.
- Emergency-stop functionality must be tested before each activity, and the results recorded.

5. Software, AI Policies and Code-Review Red Lines
5.1 Prohibited practices
The following practices must not be used in physical human-robot sports training or competition:

1. Bypassing, disabling or dynamically increasing hardware force limits, speed limits, mechanical stops, collision detection or emergency-stop protection;
2. Using undeclared debug interfaces, hidden commands, backdoors, dynamic libraries or privilege escalation to change safety parameters;
3. Switching to an unapproved high-hazard policy when losing a score, receiving a referee penalty, detecting athlete fatigue or identifying an opponent as human;
4. Connecting automated attacks, interceptions or counter-attacks directly to actuators without an independent safety controller;
5. Continuing high-risk actions when perception is uncertain, vision is occluded, sensors conflict or communications are abnormal;
6. Allowing reward functions, online learning or automatic policy updates to increase contact force, speed or attack permissions during formal operation;
7. Transferring high-risk strategies learned in robot-to-robot contests into human-robot mode without reassessment;
8. Concealing safety-relevant software versions, model changes, fault records or known defects.
5.2 Human-robot safety controller
Physical human-robot systems should use an independent safety controller or an equivalently validated architecture to constrain AI outputs. At minimum, it should check action permissions, speed/torque limits, exclusion zones, contact mode, human distance, sensor health and stop conditions. The safety controller must have final authority to veto hazardous actions.
5.3 Action prediction and response speed
Multi-camera vision, skeletal pose estimation, temporal models and action prediction may support training analysis, risk warnings and constrained action planning. Prediction accuracy or low latency must not be used as justification for greater impact. In human-robot mode, validated speed limits, distance control, compliant control and action-permission mechanisms must prevent hazardous contact.

This document does not prescribe one universal artificial-delay value for every sport. Any latency threshold must be determined through task-specific assessment of sensors, actuators, robot dynamics, contact tasks and validation results. Artificial delay cannot replace force limits, separation or emergency stops.
5.4 Code review and black-box validation
The following pre-event review process is recommended:

- Step 1: Configuration and static review. Review the hardware abstraction layer, control-loop interfaces, safety parameters, libraries, permissions, logs, version signatures and safety-related branches.
- Step 2: Simulation and fault injection. Test sensor loss, occlusion, delay, communication loss, target misclassification, abnormal motion, emergency-stop activation, actuator saturation and power faults in a digital twin or simulation environment.
- Step 3: Controlled physical testing. In an isolated facility, use test dummies or instrumented targets to validate speed, force/torque, stopping distance, repeatability and fail-safe states.
- Step 4: Independent human review. Qualified reviewers without material conflicts of interest should review critical findings and issue a written report.
- Step 5: Version lock. Record the approved hardware, firmware, model, parameters and tool configuration with version identifiers and integrity hashes. Any safety-relevant change triggers reassessment.

Static scanning or AI-assisted review alone cannot prove safety. Where a developer cannot disclose source code because of trade-secret concerns, alternatives may include confidentiality agreements, controlled on-site review, third-party audit or black-box interface testing. Sufficient evidence must still be provided to demonstrate that safety functions work. Confidentiality arrangements must not prevent incident investigation, legally required disclosure or information requests by competent authorities.

6. Perception, Communications and Fail-Safe Behavior
6.1 Multi-sensor and vision systems
Multi-camera vision may be combined with joint encoders, inertial sensors, force/torque sensors, tactile sensors and safety scanners. Camera count, frame rate, localization accuracy and end-to-end latency should be selected according to task risk and validation results. A fixed camera count or single accuracy figure should not be treated as the mandatory answer for every setting.
6.2 Failure response
If any of the following occurs, the system must enter a predefined safe state, stop or limit hazardous motion, and require a responsible person to determine whether reset is permitted:

- Critical sensor data is lost, frozen, conflicting or outside its valid range;
- Vision occlusion prevents reliable estimation of human-robot distance or posture;
- Communication delay, packet loss or clock desynchronization exceeds validated tolerance;
- A person enters an exclusion zone, the robot leaves its operating boundary or localization confidence is insufficient;
- A controller fault, watchdog event, software crash or safety-parameter integrity failure occurs;
- Overtemperature, low voltage, battery abnormality, actuator overload or mechanical damage is detected;
- The emergency-stop path or safety controller cannot be confirmed operational.
6.3 Cybersecurity and updates
- Segment control networks from public networks or use an equivalent risk-assessed network architecture.
- Disable unnecessary remote-control functions, debug ports and default credentials in production.
- Apply version control, integrity verification and change logging to firmware, models, configuration files and dependencies.
- Do not update models or safety parameters during live operation without authorization.
- Retain safety logs sufficient for incident investigation while limiting unnecessary personal-data collection.

7. Specific Rules for Human-Robot Sparring, Robot-to-Robot Sparring, Human-Robot PK and Robot-to-Robot PK
7.1 Human-robot sparring
- Start with static demonstrations, non-contact tracking, low-speed target holding and fixed trajectories.
- Early training should constrain the robot’s attack initiative and use predefined action sets, speed ranges and stop conditions.
- Having a human strike a robot-held target may be easier to control than having a robot strike a person, but risks from reaction forces, instability, target detachment, athlete hand/foot injury and robot falls must still be assessed.
- Robot movement toward or contact with a person requires a separate assessment and explicit authorization. The word “sparring” does not reduce the required level of safety.
7.2 Robot-to-robot sparring and PK
- Train in simulation before conducting physical testing in an enclosed facility.
- Physical R2R testing requires personnel exclusion zones, robot operating boundaries, remote emergency stops, protection against flying objects and a defined recovery procedure.
- Unrestricted full-power or zero-latency operation must not be assumed to be permissible. Any output limit must be validated against robot structure, contact materials, collision energy and facility conditions.
- Spectators and staff must remain outside areas exposed to falls, swinging limbs or launched equipment.
- Strategies developed in R2R contests must not be transferred to human-robot mode without dedicated validation.
7.3 Human-robot PK
A participant’s signed waiver is not, by itself, a sufficient basis for authorizing physical human-robot competition. Risk assessment, hardware and software validation, staff training, insurance and legal review, emergency-stop drills and on-site emergency readiness are required. If impact, crushing, falling or autonomous-motion risks cannot be reliably controlled, use simulation, digital contests, non-contact testing or R2R modes instead.
7.4 Athletes and minors
Participants must receive clear and understandable risk information and be able to withdraw at any time. Where minors participate, comply with local laws on guardian consent, child protection, educational institutions and sport activities, and apply more conservative controls. Consent forms do not replace engineering safeguards.

8. Sport-Specific Guidance
8.1 Taekwondo robots
- Define robot-held target training, athlete kicks to robot targets, robot demonstrations, controlled-contact human-robot training and R2R demonstrations as separate modes.
- For robot kicks, assess foot mass, end-effector velocity, hip and torso inertia, contact location, athlete balance and the recovery trajectory.
- Head, neck, spine, groin and joint areas should be exclusion or strictly controlled zones. Do not rely only on a vision-based geofence; combine it with hardware limits, action permissions and stop mechanisms.
- Taekwondo protective equipment and competition rules do not certify robot impact forces as safe.
8.2 Boxing robots
- Prefer configurations in which the robot holds mitts and the athlete strikes the targets.
- Before a robot actively punches, validate glove/end-effector padding, wrist and elbow mechanics, punch trajectory, contact force, stopping distance and repeated-strike risks.
- Do not use a human head as a high-power robot-strike target. Physical robot strikes to a human head must not occur without independent validation and formal authorization.
- Slips, counters, combinations and predictive policies must be constrained by the safety controller. Prediction of human telegraphing must never bypass action permissions.
8.3 Kendo and safety-adapted training with traditional weapons
- Use only training implements approved by the project safety lead. Inspect for cracks, burrs, loose parts, sharp tips and insecure connections.
- Define distance, speed, permitted target zones and stop conditions. Do not substitute real blades or points for training implements.
- Do not enable high-speed, rigid or uncontrolled automated interception in environments where people may be exposed.
- Concepts such as sen-no-sen, sen-no-sen interception and defensive counter-timing may be used for technical analysis and simulation, but must not justify faster physical attacks, bypassing the safety layer or automated targeting of high-risk body areas.
- R2R equipment contests must also assess breakage, flying fragments, rebound and accidental entry by personnel.
8.4 Lacrosse launching, passing and catching
- Clearly label the ball type, available speed range, launch direction, maximum operating area and emergency-stop method for each launcher.
- Physically separate launch and receiving areas. No person may stand in an uncontrolled launch path.
- Increase speed, spin, launch angle and cadence progressively from lower-risk settings, and validate each setting through measurement.
- For catching and passing robots, limit stick movement range and speed, and assess stick-head impact, ball rebound, crushing and robot-movement risks.
- If target recognition fails, a person enters the receiving zone, a ball jams or launch status becomes abnormal, the launcher must stop and require manual reset.
8.5 Cricket bowling machines
- Mark the ball type, speed range, launch angle, direction and safe operator position.
- Determine speed settings from manufacturer data, measured ball speed, ball properties, batter skill, protective equipment and facility conditions. A user-interface setting name is not a substitute for actual speed measurement.
- Provide a pre-launch warning, accessible emergency stop, launch-zone separation, jam-clearing procedure and controls against unintended repeated launches.
- Before changing the direction, clearing a jam, repairing the machine or entering the launch area, follow energy isolation and prevention-of-unexpected-startup procedures.
- Do not launch high-speed balls toward people, spectators, glass or other vulnerable targets unless the specific use has been validated and appropriately controlled.

9. Facility, Personnel, Operations and Emergency Management
9.1 On-site responsibility
Each activity must appoint an on-site safety lead and define the responsibilities of operators, coaches, referees, maintenance staff and first-aid personnel. Operators must not be assigned incompatible duties that prevent them from monitoring high-risk tasks.
9.2 Pre-activity checklist
Before each activity, confirm at minimum that:
- The risk assessment, mode authorization and operating procedure are current;
- The robot, end effectors, guards, fasteners and batteries show no abnormal condition;
- Force limits, speed limits, exclusion zones, collision detection and safety-controller settings are correct;
- Emergency-stop and safe-stop functions pass testing;
- Facility boundaries, personnel exclusion zones and projectile protection are in place;
- Sensors, clock synchronization, communications and logging are working;
- Operators are trained and participants have received risk information;
- Emergency contacts, first-aid supplies, evacuation routes and incident-reporting procedures are clear.
9.3 Incidents and near misses
If injury, uncontrolled movement, unintended contact, emergency-stop failure, bypassed limits, abnormal falls, unintended ball launches or another potentially serious event occurs, stop the activity immediately, protect the scene, provide necessary first aid and report the event as required by applicable law and event rules. Preserve necessary logs, version information, video and equipment state while handling personal data lawfully. Do not resume the same high-risk operation without investigation and written authorization.

10. Data Protection, Privacy, Intellectual Property and Transparency

- Clearly inform participants what video, movement, audio, identity, health or biometric data are collected, why they are collected, who can access them, how long they are retained and how they can be deleted.
- Collect only data necessary for training, safety, officiating or research. Do not assume permission to use participant data for advertising, model training or other secondary purposes.
- Apply additional legal review to minors’ data, health data, biometric data and identifiable video.
- Protect developer trade secrets through access controls, confidentiality agreements, isolated environments and minimum necessary disclosure. Trade-secret protection does not remove duties to provide safety evidence, support incident investigations or comply with legally required disclosure.
- Claims such as “burn after reading,” “no upload” or “no retention” must match the actual architecture, logs, backups, telemetry and third-party services, and must be verifiable.

11. Testing, Validation and Records

Every authorized system should maintain a traceable technical file containing at least:

1. Robot model, serial number, end effectors and hardware configuration;
2. Firmware, AI model, control policy, configuration parameters and dependency versions;
3. Risk assessment, hazardous-scenario list and risk controls;
4. Test results for force limits, speed limits, stopping distance, emergency stop, collision detection and fail-safe behavior;
5. Records of sensor loss, vision occlusion, communication faults, power faults and fault-injection tests;
6. Operating procedures, training records, maintenance records and pre-activity inspections;
7. Incidents, near misses, corrective actions and retest results;
8. Reviewers, approvers, dates, test conditions and calibration information.

Testing should use loads, speeds, equipment, surfaces, lighting, networks and personnel configurations representative of actual operation. Simulation results must state their limitations and cannot replace necessary physical tests.

12. Proposed WAISF Tiered Admission and Certification Framework

WAISF recommends that sport governing bodies and partner organizations progressively consider the following tiers:

- Level 0 — Simulation Only: Simulation only; no physical human contact permitted.
- Level 1 — Isolated Robot Operation: Single-robot or R2R operation in an enclosed area with personnel isolated.
- Level 2 — Non-contact Human Interaction: Non-contact human-robot training.
- Level 3 — Controlled Equipment Contact: Human strikes to robot-held targets or controlled ball interactions, subject to task-specific validation.
- Level 4 — Controlled Human-Robot Contact: Robot contact with a person may occur under controlled conditions, subject to independent safety assessment, authorization and continuous supervision.
- Level 5 — Human-Robot Competition: Physical human-robot contest or competition; consider authorization only after dedicated rules, legal review, independent validation, insurance and emergency safeguards are in place.

A tier describes the permitted operating scenario; it is not a universal guarantee of safety for every device or participant. Material changes to hardware, firmware, models, end effectors, speed ranges or operating environments must trigger reassessment.

13. Applicable Law and Disclaimer

This document is industry guidance. It does not replace the laws of any country or region, product-safety regulations, machinery-safety requirements, occupational health and safety duties, sport governing-body rules, privacy and data-protection requirements, child-protection requirements, insurance conditions or facility permits.

Before adopting this document, organizations should obtain review from legal counsel familiar with the relevant jurisdictions and qualified robotics or machinery-safety professionals. Where laws differ, binding laws and competent-authority requirements prevail. This document must not be used to reduce a statutory safety duty.

Publication by WAISF does not automatically certify, endorse or guarantee the safety of any particular robot, manufacturer, software, event or training mode. Any certification program should separately publish its audit scope, evidence requirements, independence safeguards, validity period, suspension and revocation process, and appeals procedure.

14. Implementation Recommendations

WAISF recommends that sport governing bodies, clubs and research organizations implement the guidance in the following order:

1. Begin with simulation, motion analysis and non-contact training;
2. Progress to enclosed, low-speed, low-energy testing with instrumented targets;
3. Validate safety controls, emergency stops, failure responses and personnel separation;
4. Obtain independent review and written approval of the test results;
5. Pilot the system with defined equipment, actions, personnel and venue limits;
6. Update the guidance continuously in response to incidents, near misses and new evidence.

WAISF’s core position: Progress in embodied-AI sports must advance alongside verifiable risk control. No competitive objective may take precedence over human life, bodily safety or meaningful human oversight.