World’s First Autonomous BJJ Humanoid

By Dorian Cartwright | TrustRobotics™

Building a humanoid that understands the roll

TrustRobotics is developing an autonomous Brazilian jiu-jitsu humanoid architecture that perceives a training partner, interprets the physical relationship between their bodies, selects sport-appropriate actions, and adapts throughout an exchange.

Autonomy means selecting physical responses without contemporaneous human commands choosing each technique. Instructors can establish objectives and stop the session. This is a development goal and proposed architecture, not a report of a demonstrated system or verified world-first status.

ARCHITECTURE GUIDE & TAP HANDLING →

I. Real World Perception: Sensor Fusion for BJJ Pose Estimation

Video, tactile contacts, joint encoders, and inertial measurement units (IMUs) are fused to estimate body position, orientation, and motion. The perception state also considers center of mass, inertia, momentum, detected resistance, and grip position and strength, distinguishing direct measurements from model-based estimates and retaining uncertainty.

DETAILED VERSION: REAL WORLD PERCEPTION →

II. Real World Action: Inter-Body Artificial Intelligence

Inter-body artificial intelligence interprets grips, frames, support, loading, and constraints to evaluate available actions. The BJJ LLM trained in section III uses this relational and physical state, training objectives, and interaction history to select an intended response and output a structured BJJ action token. Tokens form an action timeline for downstream validation and motion control; the physical exchange determines the actual next state.

A deep armbar attempt illustrates this approach: current body configuration, contact evidence, and interaction history inform estimates of possible outcomes. The detailed version explains the state diagram and why its illustrative probabilities depend on context.

DETAILED VERSION: INTER-BODY ARTIFICIAL INTELLIGENCE →

III. Teaching a robot a jiu-jitsu lineage

Section III trains and adapts the BJJ LLM to an instructor’s jiu-jitsu persona. Demonstrations, corrections, and supervised exchanges connect physical and relational states with preferred actions, positional priorities, teaching objectives, and responses to resistance. Section II uses the resulting model during the roll to output BJJ action tokens.

DETAILED VERSION: INSTRUCTOR PERSONA & LINEAGE →

IV. A robot must also know when to tap

Tap recognition and other termination conditions can preempt an executing action timeline. The robot must also recognize when to submit, release control, or disengage. Training preferences and predicted success remain subject to execution constraints.

Explore the architecture

ARCHITECTURE GUIDE & TAP HANDLING →

Use the detailed-version link in each section above to explore that topic. The architecture guide connects the topics through the control loop, tap handling, and technical development targets.

Read the original article on the TrustBot Blog