R4T15
Frozen R4T15 policy for the Microduck Arena sprint-2m challenge.
The original observation normalizer is baked into policy.onnx; feed raw
61-number Arena observations. The network outputs 14 joint actions at 50 Hz.
Artifact identity
- Training seed: 3.
- Selected checkpoint:
model_1249.pt, 1,250 completed PPO updates. - Original full-run recipe: 4,096 environments × 24 rollout steps × 3,000 updates.
- Training source commit:
2dc9905db36864301558cea8df7120c2109348d9. - Policy SHA-256:
db491cedf1785370c67254088e394f59410422f0455040d3bba5b53a557f947f. - Source checkpoint SHA-256:
d5ce94c2ba91f35d6d9da1e92c2243b3638507626abe702eed3358fc706055fd. - Lockfile SHA-256:
fcbf04e1fbea2cccdc81fac521a5890ea95d245227515f09b05e141b89c5b7e2. - Tracking reward weight: 4; tracking standard deviation: sqrt(0.15).
- The ONNX bytes are the frozen original, without gain/bias search or further PPO updates.
Local evaluation
Pinned Arena 0.5.1, runner 5: 2.093 seconds, 5/5 finishes, command 1.20 m/s, using the native sweep and scoring procedure. This is a local measurement, not a claim about the public Arena's result or leaderboard rank.
Provenance
The historical training checkout was /home/tamil/MD/b5/RWD4/code at the commit
above, recorded clean. This local path is preserved as provenance; it is not a
public repository URL. No public fork or independently reproducible source URL
is claimed by this publication. Exact training metadata is in manifest.json.
Original training command:
train Mjlab-Sprint2m-MicroDuck --gpu-ids 0 --env.scene.num-envs 4096 --agent.max-iterations 3000 --agent.seed 3 --agent.logger tensorboard --agent.run-name R4T15_s3
This repository publishes the selected ONNX policy and its manifest. It does not include training logs, optimizer checkpoints, credentials or other runs.
- Downloads last month
- 11