Three arms
Same rider, same bike, same controls, same shoves. Only the brain changes.
- Think: Claude Code controls every 20 Hz step through an MCP server. Sim pauses while it thinks. No code, no memory between rides.
- Practice: Claude Code writes the controller and practices as many times as it needs. Basically LLM-driven exploration. Scored after attempt 1, at checkpoints and at submission.
- Train: an RL policy (PPO) trained with randomized shoves, backpack and friction.
Connections
Parent: Thinking Faster
Tangents
- RAX: Arm 2 is RAX on a dynamic task