craigm26/microduck-duckbatch-b002-256x128x64

A walking policy for Pollen Robotics' Microduck: 57,934 parameters (61-256-128-64-14), distilled from Pollen's default walker (velstand.onnx, 197,774 parameters) by duckbatch, batch b002-student-size-longer, attempt a01.

Drive it in your browser (Pollen's simulator loads it from this repo).

Simulation only. It has never run on a real robot. A student that matches the teacher in mjlab is a candidate for a hardware test, not a result about hardware.

Measured

mjlab Mjlab-VelStand-Flat-MicroDuck with domain randomization, stumble pushes and observation noise on; the task's deliberate prone spawns and topple pushes are off for the walking numbers and on (every episode prone) for the get-up numbers. 30 s per seed on held-out seeds 2001, 2002, 2003; mean ± spread over seeds. The teacher ran in the same eval.

This policy Teacher
Falls per minute 0.38 ± 0.07 0.14 ± 0.02
Time fallen 0.85% ± 0.07% 0.43% ± 0.10%
Planar velocity error (m/s) 0.178 ± 0.004 0.174 ± 0.004
Yaw rate error (rad/s) 0.388 ± 0.007 0.368 ± 0.003
Gets up from prone within 6 s 98.2% ± 0.5% 96.9% ± 1.5%
Time to get up (s) 0.56 ± 0.01 0.56 ± 0.01
Parameters 57,934 197,774
FLOPs per step 114,944 393,728
1-thread latency, p50 (x86 laptop) 14.5 us 29.0 us

Use

  • Browser: https://pollen-robotics-microduck-simulator.hf.space/?move=craigm26/microduck-duckbatch-b002-256x128x64
  • Robot: robotctl policy load walk craigm26/microduck-duckbatch-b002-256x128x64 (manifest schema 2, walk slot)
  • Contract: obs[1,61] -> actions[1,14], the normalizer baked in, the same as every Pollen policy.

Full record (training trace, every judge decision, the decision-model answers): https://github.com/craigm26/duckbatch, records/b002-student-size-longer/.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading