tags:
- microduck
- robotics
- onnx
- microduck-slot:walk
library_name: microduck
pipeline_tag: robotics
dfreq-rollers — N8 coupled feedback update
Frozen Roller policy at external command 1.1. Local independent evaluation: 128/128 finishes, mean 2.7175 s, median 2.720 s, maximum 2.760 s. The actual official local CLI at command 1.1 returned 2.715 s, 5/5 finishes on its prescribed five seeds. These are local measurements, not a new Arena server result or leaderboard claim.
Method and contract
The original credited Pollen neural backbone remains frozen. N8 adds a memoryless coupled feedback head using 16 bounded projections of existing hidden features plus a bias, with 238 learned coefficients. Episodic derivative-free optimization and accepted-direction search produced the selected model; this is not PPO. Its total added residual is bounded by 0.12 rad. Actions are radian offsets from DEFAULT_POSE; that residual cap is an architectural choice, not an official whole-action limit. The inherited feedback command calibration is retained.
The deployed model receives only the official 61 raw feedback inputs and emits 14 float32 actions at 50 Hz. No scoring clock, seed identity, reset fingerprint, privileged simulator velocity/position or per-seed action table enters the policy. Previous-action feedback remains the actual previous output. Official physics, preparation and actuator limits were unchanged.
Frozen evaluation
N8 and command 1.1 were frozen before one evaluation on 128 previously untouched paired resets (900001–900128). No alternative policy or command was selected from those results. N8 was faster on every pair against both the earlier private fallback and previous published policy. The private fallback also finished 128/128, mean 2.922617 s: N8's paired mean gain was 205.117 ms (paired bootstrap 95% interval 199.883–210.274 ms). The previous public model finished 128/128, mean 3.178906 s; its 10.755 s recovery outlier affects that mean. Median paired gain against it was 390 ms. All individual results, failures from training and fixed promotion gates were retained privately.
The original prescribed five seeds were already exposed regression data. Their CLI replication is not fresh validation. The 128 final starts are limited simulation evidence, not a guarantee on all resets, other physics or hardware. Floating-point contact sensitivity limits portability; claims here concern the exported ONNX, not Python training inference. Later private experiments did not replace this validated N8 model.
Runtime pins: Arena source bbddfe587d300063b1d162281d647740738742f5, runner v5, Python 3.12.12, NumPy 2.4.1, MuJoCo 3.10.0, ONNX Runtime 1.24.4 on Apple Silicon macOS. Preparation is 25 untimed control calls; distance origin resets afterward while momentum is retained.
Artifact, history and credit
policy.onnx SHA256: dc660e8128c3ad2446a0cee9c2f8690ccae3dab681509c57010f91a59d78e797.
Previous public revision: 6f282e1a904146ef800ef011b02863c29e2a420b, model SHA256 58122a7287b3330d167f5659a5a8fc9f65c7456f7851e5e5d0a9673ecc505d92. That previous entry's historical server score was 3.081 s; publication of N8 alone does not submit a new Arena entry.
Original neural network: Pollen Robotics Roller Sprint reference, revision c72fb92594e8b415487e87dfeed509141afc5e25, SHA256 01edf63a529a8396d46dac62bb427fda1ce67685611e72ddf790bf46f610c82a. The manifest training field describes upstream Pollen training. Existing service-scoped remix provenance is preserved. No standalone open-source model license or blanket off-platform redistribution grant is asserted.
Simulation only; no real-robot deployment or braking/stability safety claim. This publication changes only the Roller repository, with owner approval.