Sprint forward-priority PPO, seed 1
PPO actor-critic with shipped actor mirror-consistency loss coefficient0.5 (no trajectory augmentation), forward tracking reward2 to4, and constant2% standing environments rather than ramping to25%. Combined intervention; no causal attribution or state-of-the-art claim.
Trained from scratch:4096 environments,3000 iterations,seed1,local RTX4090. Selected checkpoint750. Actor512/256/128 ELU MLP;61 observations to14 actions at50Hz, normalizer baked into ONNX. Uses forward twist command; head/body commands zero for competition.
Local simulation before publication:2.767s5/5 at0.80m/s versus baseline5.123s5/5. Paired additional starts6-25:2.7535s20/20 versus baseline5.11775s20/20, each at previously selected command. Local results are not official Arena results. One training seed; no physical tests or Hubot transfer established.
Reproduce
Base repository https://github.com/pollen-robotics/microduck-challenges at4b675ebf4d7c7b961cd78da94bfc37ca1686d321. Local commit5518dc485740073aa711270ee4ce379b7feedb47 has not been pushed upstream. Manifest repo records original provenance; the local commit is not available from Pollen. Apply included recipe.patch to the pinned base:
git clone https://github.com/pollen-robotics/microduck-challenges.git
cd microduck-challenges
git checkout 4b675ebf4d7c7b961cd78da94bfc37ca1686d321
git apply /path/to/recipe.patch
uv sync --group workshop
uv run train Mjlab-Sprint2m-MicroDuck --env.scene.num-envs 4096 --agent.max-iterations 3000 --agent.seed 1 --agent.logger tensorboard --agent.run-name sprint-forward-sym-s1
GPU training is not bit-reproducible. Included checkpoint and ONNX identify this candidate.
Checkpoint SHA256: 2e04836b4ad840265e1423130c48f8a5a344cca69f623a8658b043bd9894e6dd
ONNX SHA256: a91734416efb84ef1df92a6f38db168b5ea9adac72395f387a9f70db453d09b2
- Downloads last month
- 18