Tromset/flymotion-commands

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": "{\"safe\": true, \"steps\": [{\"action\": \"turn\", \"direction\": \"right\", \"value\": 90}, {\"action\": \"walk\", \"direction\": \"forward\", \"value\": 3}], \"repeat\": 1}"}
], "lang": "fr"}

• 3,000 examples (2,700 train, 300 validation), half French, half English.
• Actions, values, order: safe (false for a refusal, with no steps), then each step: an action (walk, turn, circle, look, wait, stop or a skill), a direction in words and a positive value. Nothing else: the FlyMotion bridge writes the sentence for the user and turns the plan 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 as actions (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, compiles for the Microduck profile, and walks for real with the gait of Pollen's simulator (no time stepping in place). 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.