README.md

metadata

license: mit
language:
  - fr
  - en
task_categories:
  - text-generation
tags:
  - robotics
  - motion-planning
  - microduck
  - function-calling
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl
      - split: validation
        path: data/valid.jsonl

FlyMotion commands

French and English requests for a small walking robot (Microduck by Pollen Robotics), each paired with the motion plan the robot should run. It trains and evaluates the FlyMotion planners (GitHub, Ollama).

{"messages": [
  {"role": "system", "content": "You are FlyMotion, the motion planner of a small robot..."},
  {"role": "user", "content": "Tourne à droite, ensuite avance 300 cm"},
  {"role": "assistant", "content": "{\"say\": \"Je tourne de 90 degrés vers la droite, puis j'avance de 3 m.\", \"repeat\": 1, \"steps\": [{\"action\": \"turn\", \"direction\": \"right\", \"degrees\": 90}, {\"action\": \"walk\", \"direction\": \"forward\", \"meters\": 3}]}"}
], "lang": "fr"}

• 3,000 examples (2,700 train, 300 validation), half French, half English.
• Plans in words: directions are forward, backward, left, right; distances and angles are positive. The FlyMotion bridge turns them into velocities, durations and limits from the robot profile, so a model never handles signs or physics.
• Covered: walking in four directions, turning, circles, polygons with repeat, head movements, the six Microduck skills (kick_left, kick_right, roulade, ground_pick, sit_toggle, polite-bow), waits, stops, refusals of impossible or harmful requests (jump, fly, kick the dog), and non-motion questions.
• Checked: every plan passes the validator and compiles for the Microduck profile. None of the 32 benchmark requests appears in the data, so the benchmark stays a fair test after training.

Generated by training/build_dataset.py (deterministic seed) and republished from GitHub on every change.