YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
dfreq-trained-b-mps-early
Private trained candidate for roller-sprint-2m, roller accessories, 61float32inputs and14float32 radian-offset actions at50Hz. No publication performed.
Native3 CPU MuJoCo/BAM rollouts plus MPS clipped PPO; 1806 neural output-layer parameters trained, all earlier weights and inherited feedback constants frozen.
Local unchanged native evaluator: official sweep 2.716s,5/5finishes at command1.1. Fresh paired holdout128/128finishes, mean2.715273s versus generalbaseline2.717148s. Paired gain95%interval[-0.001757812499999963, 0.0053906250000000065]. Do not claim superiority when this interval crosseszero. Both32seed stresscommandsets1.0 and1.2 finished completely.
Existing published2.570sbenchmark specialist remains unchanged and is faster on its exposed prescribed starts. These are general trained daily-run candidates, not a claim of improved leaderboard score or points awarded. Current timed best should be retained; no Tugcup latest-entry change.
No new seed,time,position,fingerprint or per-seed lookup enters these policies. Previous-action feedback is the actual previous output. Input/output contract, native physics and actuator limits unchanged. Original inherited policy lineage/terms are in upstream-base-card.md. New calibration/PPO work does not relicense inherited weights.
SHA256 efae4a9b3516b2d0ebf9958f2678e7d47357612784d9bfa3fb001ff16c66ac7d. Training/selection frozen before fresh128seed evaluation. Full evidence is in EVALUATION.json.