Microduck Roller PPO checkpoint 64
Genuine independently updated PPO checkpoint from our native CPU physics/MPS learner. Legal 61-observation to 14-action controller; no seed lookup or privileged inputs. Event: roller-sprint-2m, rollers, command speed 1.0 m/s.
Local original-evaluator score at command 1.0 m/s: 2.720 seconds, 5/5 finishes. Separate validation: 128/128 complete starts and 64/64 command-stress starts. This is a daily-run candidate, not a personal-best or podium claim. Existing published best is 2.570 seconds. Timed leaderboard retains that best; this model does not affect the Tug cup entrant.
Training checkpoint 64, not a hash-only variant. No third-party opponent weights or outputs were copied. ONNX SHA256: 4deaf32f4d8318ee21e639fa7038da2ef3afd3c7707bdb1ed51870167ebdf31b.