How Humanoid Robot Fighting Actually Works

A movement still needs balance

A fighting humanoid coordinates arms, legs and torso while its feet support the body. Joint controllers execute movement commands, and balance software reacts to measured motion. A person, a stored sequence or a sensor-based policy can choose the action. The visible movement alone does not identify that choice.

Who chooses the movement?

[1]

These descriptions can overlap. A person may select a motion derived from motion capture while a learned tracking policy controls the joints and balance. [2]

The distinction between operator commands and local motor control is developed in the teleoperation guide. A balance loop can run locally while the person still chooses every action.

Torque is not a measured impact force

Shoulder and elbow joints position the arm. Hip, knee and ankle joints support the body and reposition the feet. A waist joint, where fitted, changes torso orientation. Electric motors produce torque, measured in newton-metres. Reduction gears raise output torque while reducing speed. Friction reduces transmission efficiency, while gearing increases the rotor inertia felt at the joint. A quoted peak joint torque describes rotational effort at that joint. [3]

The joint-torque contribution needed to balance a hand load obeys τ = JᵀF. J describes the arm’s current geometry, and F contains the balancing force and moment at the hand, expressed in the same frame. The motors must also support the arm’s weight and overcome other loads. During a collision, inertia and contact duration matter. Dividing a torque specification by an arm length cannot establish measured striking force. [4]

Holding a pose against contact and accelerating into free space place different demands on the same motor. A permitted trajectory must stay within position, speed and torque limits throughout its motion. Contact can invalidate the original timing before the arm reaches its target. This is why a motion controller needs a way to slow, stop or change the reference. [5]

Why the torso and legs move too

Moving an arm changes the distribution of mass and angular momentum. Lifting a leg removes one ground contact. The remaining foot must support the body while the torso and other limbs move. Whole-body control can assign priorities to support, posture and limb tracking. A balance correction can interrupt the intended arm motion when the two tasks conflict. The support foot must also avoid sliding and tipping. Controllers can constrain the ground force using available friction and the position of pressure under the sole. [5]

The relationship between foot contact, pressure and body acceleration is explained in how humanoid robots walk without falling. The same physical constraints remain during robot sport.

An IMU measures body rotation and acceleration-related signals. Encoders report joint motion. Dedicated force-torque sensors can measure contact loads; some systems estimate external loads instead. Seoul National University researchers tested such estimation on the TOCABI humanoid using joint and inertial measurements, with foot sensors providing comparison data. Their published test concerns walking, so it does not validate impact detection during a match. [6]

A prepared trajectory can coexist with closed-loop control. The decisive question is which measurements change which command. Does the controller adjust an ankle after body rotation, stop an arm after contact, or select another action from an opponent’s position? These are separate claims. A sequence that looks responsive on video cannot establish all three. [7]

An actual rulebook changes the task

The World Robot League organizer lists humanoid battles at its Suwon festival on January 16–18, 2026. Its public Humanoid Battle rulebook, revision 25.06.30, sets out standard one-on-one bouts for small competition robots and separate provisions for group rumble events. These rules should not be assigned to full-size T800 matches. [8] [9]

[9]

Equipment rules require inspection and prohibit sharp parts and devices intended to cause excessive damage. Batteries must be secured in the upper body and protected by a cover. Swollen batteries are prohibited. These are event-specific controls on hardware and handling, alongside the referee’s authority during a bout. [10]

What EngineAI’s published tests establish

EngineAI’s URKL rulebook, dated March 2, 2026, describes a separate technical assessment. Teams submit T800 motion policies and simulation material, including motion-capture processing. Physical tests include recovery inside a circle 8 m in diameter. The recovery scale awards 15 points for 0–3 s, 10 for 3–10 s and 5 for 10–20 s. These are assessment bands, not a claim that every T800 meets them. [1]

The rulebook also describes strikes measured on calibrated equipment. Its main-tournament format section still awaits an announcement. A qualifying motion score therefore cannot establish the scoring rules of a head-to-head bout. [1]

Variant-specific hardware and the limits of the published evidence are covered in the EngineAI T800 technical analysis.

What a complete match report should show

Force limitation starts with controlling drive effort. For example, the ROBOTIS XM430-W350 manual describes a current limit in torque and current-based position modes. It also documents temperature monitoring and configurable protective shutdown. These controls protect the actuator within their stated modes. They cannot, on their own, certify the force of a collision at a robot’s hand. [11]

A robot that gets up has recovered a stance. That does not establish an undamaged joint, battery or sensor. The RSS 2025 G1 getting-up research tested recovery from face-up and face-down starts on several surfaces. It was not a test of repeated blows or accumulated match damage. [12]

Useful competition reporting identifies the controller, operator role, robot variant, surface, allowed contact and stoppages. It should count failed actions as well as successful ones. Contact loads, actuator temperatures, recovery attempts and operation after the match help assess repeated-load durability. Without those records or a disclosed control system, a short sequence cannot establish durability or autonomous opponent tracking.