binhpham/reachy-mini-massive-motion-library

Reachy Mini Massive Motion Library

10,872 episodes (15.1 h at 25 fps). Each one is a prompt performed by Reachy Mini: the teacher's motion recipe, turned into 25 Hz motion by the plan-to-motion flow generator, projected onto the robot's reachable set, and rendered in MuJoCo with the official model.

source	episodes	what
astra	5,000	5,000 individually authored scenarios (Astra)
astra_lively	5,000	the same 5,000 prompts re-authored in the lively house style (Codex, gpt-6-astra)
claude	520	hand-written by Claude (Anthropic)
claude_events	65	build-up/release events and long multi-phase stories (Claude)
seed	287	the original 287 hand-written recipes

• observation.state / action: the 9-DoF target trajectory [x y z roll pitch yaw antenna_right antenna_left body_yaw] (metres / radians). Play it with the SDK by converting to a recorded move.
• task: the prompt. meta/teacher_rows.jsonl gives each episode's source, family, idea and recipe.
• astra and astra_lively contain the same 5,000 prompts performed in two styles: the original choreography, and a re-authored "lively" version with more energy, rhythm and ear expression.

Known issue: in 29 of the 10,872 episodes the video encoder dropped one frame, so observation.images.front is one frame shorter than observation.state / action (which are complete).

Pipeline: LLM planner -> motion recipe -> plan -> flow-matching generator -> reachability projection -> MuJoCo render.