Roller Explore032 trained research checkpoint

Native CPU physics with MPS PPO output-layer learning, checkpoint 32. The inherited Pollen neural backbone and prior own N8 feedback controller are credited in upstream-base-card.md. All 1,806 output-layer neural parameters differ from the already submitted Explore016 checkpoint. Graph topology and other weights are unchanged. This is genuine training, not a hash-only variant.

Event: roller-sprint-2m. Command: 1.1. Accessories: rollers. Float32 ONNX input obs[1,61], output actions[1,14], 50 Hz. No custom operators, external tensors, clock, seed lookup, or cached-action table was added.

Local validation

The full original Arena local speed-sweep evaluator returned 2.753 seconds, 5/5 finishes. Frozen fixed-command canonical times: 2.725, 2.730, 2.775, 2.720, 2.815 seconds. November 7 fresh validation finished 32/32 ordinary starts and 16/16 command-stress starts. Earlier validation finished 128/128 ordinary and 64/64 stress starts. No weights or command were tuned on these checks.

Limitations and purpose

Points-only candidate, not an improvement. It is slower than the published Roller personal best of 2.570 seconds. The original improvement goal and later +30 ms noninferiority guard failed and remain failed. Passing fresh finish checks does not reverse those failures. This package was prepared for a new request for distinct valid daily entries, with that speed tradeoff explicit.

These are local Apple Silicon simulator results, not official server scores. Arena acceptance, completion and points are not guaranteed. No Tug policy or Cup entry was modified. Simulation only; no real-robot safety guarantee.

Inherited terms and attribution remain in upstream-base-card.md and LICENSE-NOTE.txt. No new blanket redistribution license is asserted.

Model SHA256: 24e8339dbee01a67db643d84cd9f85d8722a8309c73550f33bdee0c10e0cf3d4.

Downloads last month

-

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