Make Reachy Mini move with emotion in real time

Binh (@pham_blnh) split the problem in two. A fine-tuned language model writes a short motion script from your prompt, and a small flow-matching model trained on real Reachy Mini clips turns it into smooth 25 Hz motion with antenna flicks and overshoot. There are three planner sizes (27B, 4B and 0.8B), a browser demo at https://huggingface.co/spaces/binhpham/reachy-mini-motion-generator, and an open dataset of 10,872 prompts performed by the robot (about 15 hours of motion). The companion reachy-animation library plays generated moves live, blended with idle breathing and head sway while the robot talks.

Hardware & cost: A Reachy Mini or its simulator. The motion server needs Linux, an NVIDIA GPU and Docker (an RTX 5090 runs the 4B model, per the README). The planners also run on a Mac or CPU, more slowly.

Difficulty: Hard.

Agent install prompt (paste into Claude Code, Cursor or Codex): "Clone https://github.com/pham-tuan-binh/reachy-motion-generator and follow its README. If I have an NVIDIA GPU, start the 4B server with docker compose and test /generate-dense with the prompt 'sneezing'. If not, use the offline pipeline with the 0.8B planner. Then play the result on my Reachy Mini (or the simulator) with reachy-animation."

Seen on X: https://x.com/pham_blnh/status/2106239250748555447

Source: pham-tuan-binh/reachy-motion-generator on GitHub: https://github.com/pham-tuan-binh/reachy-motion-generator (facts checked against the project README, Oct 2026).