GPT × Unitree R1 — Agentic Control Architecture & Workflow
Three-layer agent hierarchy (Director → Managers → Execution) with a real-time correction loop and a skill-learning loop. Beginning-to-end pipeline.
Front doorDirectorManagersExecutionSafetyRobotFeedbackLearning
A · Runtime pipeline — from prompt to action
Front door — intake & consent
1User Prompt
Natural language goal, e.g. “make coffee.”
 →
2Hardware Retrieval
Identify current Unitree R1 state (pose, battery, joints, reachability).
 →
3Event Confirmation
User approval to proceed — consent gate before any motion.
↓
Layer 1 — Director
4DIRECTOR — GPT-6
Interprets the request · runs task analysis & decomposition · decides which Managers to engage and assigns their high-level responsibilities. First checks Personal Skill Memory for a matching, reusable skill.
↓
Layer 2 — Managers (decompose responsibilities into subtasks)
5Navigation
Movement & path planning; get the base where it needs to be.
Manipulation
Arms / hands / grippers; grasp, pour, place.
Task (Beverage)
The coffee procedure: sequence & recipe logic.
Safety
Constraints, limits, collision & hazard rules for every subtask.
↓
Layer 3 — Execution agents
6Execution Agents
Translate each subtask into concrete Unitree R1 API commands — motors, joints, actuators, trajectories, timing.
↓
7Safety Validation Gate
Hard gate on every command batch: limits, collision & hazard checks. Nothing reaches the robot unvalidated.
↓
8UNITREE R1 — executes actions
Physical execution in the real world.
↓
Feedback & evaluation
9Sensor Feedback
Sensors report real-world results back up the chain.
 →
10Result Evaluation
Does the outcome match expectation?  YES → task complete.  NO → correction loop.
↻ Loop 1 — Real-time correction (fast clock)
If result ≠ expectation, evaluation feeds back to the Director/Managers for RE-PLAN → RE-EXECUTE, then back through the safety gate to the R1. Repeats until the outcome matches.
B · Skill-learning workflow (slow clock) — triggered by “watch me, learn how I do it”
DemonstrationData capture of the user performing the task
 →
Task ModelingBuild a candidate skill
 →
First ExecutionR1 attempts it
 →
User ValidationRight or wrong?
 →
CorrectionRehash / reprocess if wrong
 →
Validated SkillConfirmed & versioned
↓
★ Personal Skill Memory
Validated skills stored per user. “make coffee” becomes “make MY coffee.”
 →
↻ Loop 2 — Retrieve & adapt
On future requests, the Director retrieves the stored skill, adapts it to context, and the R1 executes; sensors re-evaluate and can trigger Loop 1 again.
Conceptual architecture — two loops on different clocks: Loop 1 corrects execution in real time; Loop 2 teaches and stores skills over time. Preview build; exportable to PDF.