sleepyduck
Yawns, sits, nods off three times, slumps and snuggles onto its cheek, asleep.
An episodic policy for the microduck (61-D observation, 14 actions, 50 Hz). Runs 10.5 s and ends asleep: sitting, slumped forward with its head on the floor, turned onto its cheek.
Run it on a robot
sudo robotctl policy add sleepyduck drfonz-indigo/microduck-sleepyduck
robotctl robot do sleepyduck
Things to know
- It finishes asleep. When the 10.5 s are up, the duck is sitting with its head on the floor, so the next policy has to be one that can get it up from there. That hand-off has not been rehearsed yet: keep a hand ready the first time.
- The beak is not part of the policy. The jaw has no servo among the 14 actions. To open the beak for the yawn and the snore, the runtime must drive it.
- Not yet tried on a real robot. Evaluated in simulation only: in 256-duck evaluations from a still stand, with the standard robot and with randomised ones, it never toppled; it sat down by about 1.8 s and stayed seated for every nod; it nodded exactly three times in 100% of episodes, its head turning a little further each time (9°, 17°, 26°); and it ended asleep and snuggled (head turned about 80°) in 100%. Its head meets the floor at 0.20 m/s (median; 0.45 m/s at the 99th percentile).
The observation normalizer is baked into policy.onnx; feed raw observations.
manifest.json follows schema 2 of the microduck policy manifest (docs/policy-manifest.md in the daemon repo).
Training
- repo:
drfonz/microduck_rl(design notes and every run:docs/SLEEPY_DUCK.md) - branch:
sleepy-duck - commit:
bf84b545d - policy: run 9, iteration 4,999 (
sleepyduck-flat-4999.onnx)
Thanks
Built on Pollen Robotics' open microduck hardware and their training code, pollen-robotics/microduck_rl.
- Downloads last month
- 22