Roller Explore040 trained research checkpoint
Native CPU physics with MPS PPO output-layer learning, checkpoint 40. The inherited Pollen neural backbone and own N8 feedback controller are credited in upstream-base-card.md. 1806 output-layer neural parameters differ from the submitted Explore016 checkpoint. Graph topology and other weights are unchanged. This is genuine training, not a hash-only variant. Neural weights also differ from Explore024, Explore032 and the other November 9 packages.
Event: roller-sprint-2m. Command: 1.1. Accessories: rollers. Float32 ONNX input obs[1,61], output actions[1,14], 50 Hz. No custom operators or external tensors. No clock, seed lookup, or cached-action table was added.
Local validation
The unchanged original Arena speed-sweep evaluator returned 2.777 seconds, 5/5 finishes. November 9 fresh validation finished 16/16 starts across commands 1.0, 1.1 and 1.2. Fixed-command canonical five-seed times: 2.755, 2.7800000000000002, 2.77, 2.775, 2.805 seconds. All 64 synthetic inference probes produced finite outputs. No weights or command were tuned on these checks.
Limitations and purpose
Points-only candidate, not an improvement claim. The historical development noninferiority goal failed; passing fresh finish checks does not reverse that failure. This package is a distinct learned checkpoint prepared for daily entries, with the speed tradeoff explicit.
These are local Apple Silicon simulator results, not official server scores. Arena acceptance, completion, duplicate eligibility and points are not guaranteed. Simulation only; no real-robot safety guarantee.
Inherited terms and attribution remain in upstream-base-card.md and LICENSE-NOTE.txt, unchanged from the prior package. No new blanket redistribution license is asserted.
Model SHA256: a6a2825af3e7ab297b484656992e174e533b5836ae0c1ebd8d8c4c7d13f64c65.