RollerPredictiveTarget
Adds a leg-only soft target-error limiter: predicted position is observed joint position plus 0.02 times the one-step-old observed joint velocity. Excess target error beyond 2.0 radians is attenuated by 0.15. Head and neck outputs remain unchanged.
512/512 independent starts finished, matching the baseline. Mean time was 2.718057 s versus 2.717930 s; the paired difference interval spans zero. Local official sweep 2.711 s versus 2.715 s. There is no established general speed improvement.
Event: roller-sprint-2m. Local evaluator command: 1.1. Accessories: ['rollers']. Float32 ONNX obs[1,61] to actions[1,14], 50 Hz, with standard operators and no external tensors. Model SHA256: d055816baf772952bfa85395a47405e9dc2f34588032e30b73ff183338fb462a.
The inherited neural model, feedback controller and contributor credits remain in upstream-base-card.md. The new branch changes actual observation-dependent actions, rather than metadata alone. No time, seed, lookup table, position, heading or privileged simulator observations are added to inference.
Scope
Experimental observation-feedback controller. This is not a replacement for existing personal-best entries, and those entries remain unchanged. Local Apple Silicon simulation cannot guarantee hosted scores, duplicate eligibility, or points. Native source revision: 9bf04c0668145528035c8a315d5d8ae2530c1279, runner version 5. Simulator and asset files were unchanged. Full upstream pytest was not run because pytest is absent in the existing interpreter; direct checks and simulator evaluations passed.
Simulation only; no physical robot testing or safety claim. Inherited rights and attribution are retained without asserting a new blanket model license. Publication and Arena evaluation do not establish physical-robot safety.
Parent-model SHA256: dc660e8128c3ad2446a0cee9c2f8690ccae3dab681509c57010f91a59d78e797.