jhp123/microduck-sprint2m-optimized

jhp123 Microduck sprint2m optimized

Own StageD forward3/progress5 training seed1 checkpoint1000. Original official CPU ONNX export with embedded observation normalization, without output gain or bias calibration. Interface:61 observations to14 actions at50Hz. No foreign policy weights, seed classifier, counter or per-seed action tables.

Result and limits

Local unchanged Arena runner5: 1.772s,5/5 finishes, command1.20m/s. Five finish times1.800/1.765/1.760/1.780/1.755s. All33 races, including the21-command scan, are retained in evaluation-local.json. Official Arena determines the submitted result.

This is a15ms (0.84%) nominal local improvement over the same run's previous official checkpoint750,1.787s5/5, which remains available at revision362a7b1a7c79e866676956eb952818746a875c98. Both are original uncalibrated exports of the same training run; only checkpoint selection differs. The small difference is not a statistically established advantage. Matched StageC training seed1 checkpoint1000 scored2.341s5/5; one matched training seed does not establish a general causal reward benefit. Another training seed has not yet replicated StageD. No extra race-seed or physical robot validation has been performed.

Earlier StageD checkpoints are retained:250=12.467s4/5;500=30.859s2/5 with three falls, each penalized50s. The two successful500 races were fast, but were not reported as the overall score.750 recovered1.787s5/5. This training run is ongoing;1000 is the best verified checkpoint so far, not a completed-run final best.

Training and provenance

• Clean source ec7027bdbcf4175677179f8562592063e292fce4. The only executable change relative to own StageC c24a50146aab950fc306e8430b85e15b1a1477e5 is forward_reward2→3. Progress5, upright2, air3, action-rate curriculum, robot, observations and physics remain unchanged.
• Progress term5*clip(actual body-x linear velocity,0,2.4)/2.4; nonfinite velocity rewarded0. Body-forward velocity is not world-lane progress.
• Run2026-10-28_09-56-42_v3_forward3_progress5_seed1_cu129; seed1, from scratch4096 environments,3000 iteration target,save250,checkpoint1000,no resume.
• RL commit8d0dbc94cbbef3f1dcae461a4dbc1a43bb9886a0. Arena commit e05128f36bc45c453509fe32fdd9614d3731025e,runner5; seeds, scan, course and failure penalties unchanged.
• runtime-cu129-v1:Python3.12.12,Torch2.9.1+cu129,CUDA12.9,A80080GB,MuJoCo3.10.0,mujoco-warp3.8.1,warp1.12.0,ORT1.24.4,NumPy2.4.1,mjlab1.3.0. This differs from the original Linux cu128 lock.
• source-identity-proof.json is a preflight-interpreter probe confirming variant import paths,clean HEAD and source/blob hashes. It is not an in-training-process inspection. Independent audit binds provenance,agent/env configuration, formal iteration logs and checkpoint manifests; original RSL Git records are retained. The shared installed RSL logger can record its v0 installation ancestor, so it is not used alone to identify the challenge source.
• Independently audited actor weights and normalization must match the original checkpoint. Normalizer is baked in by the official exporter; no static action transform is appended.
• Progress-reward concept informed by Apache2 rocPAI-Forge/microduck_rl_unilab commit02a1fbce34c2334873159e473a231fb30ee57594; independently implemented, no weights imported.

Frameworks: microduck-challenges, microduck_rl.