Teaching a Robot a Jiu-Jitsu Lineage

By Dorian Cartwright | TrustRobotics™

← EXECUTIVE SUMMARY

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.

An instructor might emphasize positional stability, guard retention, controlled transitions, or opportunities for the student to practice escapes. The architecture can represent those preferences through a methodology or persona that influences candidate-action selection.

Demonstrations, corrections, and supervised exchanges can provide examples of the instructor’s decisions under different physical conditions.

The same estimated state could therefore produce different selected actions for different training objectives. A competition-oriented policy might prioritize an advantageous transition, while a teaching policy might preserve a position long enough for a student to recognize and practice a response.

Outcome prediction and instructional preference remain distinct: the most likely successful action is not necessarily the most useful teaching action.

Training the BJJ LLM

Training examples associate the fused pose and physical state, inter-body relational state, and training objective with the instructor’s selected action and resulting outcome. Demonstrations can be annotated with BJJ action tokens, while corrections identify when the model should choose a different response. Time alignment connects the decision to the observations available when it was made.

The learned persona captures preferences such as positional stability, guard retention, controlled transitions, preferred technique sequences, and when to give a student space to practice. Supervised adaptation and instructor preference feedback are proposed ways to teach those choices. Different personas can prefer different permitted actions from the same estimated physical state.

During runtime in section II, the trained BJJ LLM outputs action tokens using the current state and selected persona. Training and model updates occur through the section III workflow; observing a new exchange does not automatically change model weights. Evaluation should check whether generated tokens reflect the intended methodology, remain grounded in the observed state, and respect execution and termination constraints.

Related architecture details

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

II. Real World Action: Inter-Body Artificial Intelligence →

ARCHITECTURE GUIDE & TAP HANDLING →

← EXECUTIVE SUMMARY