dfreq-hill

Published on Hugging Face. Repository: djduecaster/dfreq-hill; proposed Arena entry: dfreq-hill. Not submitted to Arena.

An event-specific, memoryless command calibration of Pollen Robotics’ Hill Climb reference feedback policy. The local full official sweep scored 5.848 seconds, 5/5 finishes, compared with 5.880 seconds for the unchanged reference: a modest 0.032-second (about 0.54%) improvement. This is not a leaderboard record or a server-verified result.

Credit and the actual change

The base is Pollen’s Hill Climb reference, pinned at d319208d92896353f581be754f6325d01f5e3732.

• Base SHA-256: ea590cadd5c675bc77558e57d17e43ab1b9f084449dd0e811c33077c01f38110.
• Candidate policy.onnx: 794,338 bytes; SHA-256 8fa570f375c2c399f9a044be883d0a5564d8b983ac75531f4a8679050034c635.
• All original neural weights are unchanged. No new neural training or fine-tuning occurred. The retained training metadata describes Pollen’s original model.

Only raw forward-command observation 48 is mapped, before the original normalizer:

vx_internal = Clip(float32(vx_external + float32(0.025)),
                   float32(0.15), float32(1.0))

The wrapper uses standard ONNX Split/Add/Clip/Concat operators. The other 60 observation values, including genuine actual previous-action history, are untouched. It contains no clock, step counter, seed classifier, seed table, or runtime metadata conditioning. The controller remains ordinary neural feedback rather than a prerecorded action sequence.

Interface and important scope limit

Model API 1: 61 raw float32 observations → 14 float32 actions at 50 Hz, legs rig, normalizer embedded. The event passes external [forward velocity, 0, 0]; head/body commands are unused. Kind remains perpetual with a standing entry pose.

Simulation-only and specialized to the positive Hill Climb race grid. This is not a general walker or real-robot safety recommendation. The unconditional clip also turns a zero or negative forward command into +0.15 internally. Consequently, the inherited command.idle field is not a validated stop guarantee. No robot-loading instructions or idle/braking safety claim are provided.

Local verification

The unchanged Arena runner v4 evaluated all 18 official command-grid speeds on seed 1, then completed the best three commands across seeds 1–5 (30 races total). The selected external command was 0.55 m/s, internally represented as float32 0.574999988079071 m/s.

Start seed	Finish time (s)
1	5.88
2	5.84
3	5.82
4	5.86
5	5.84

A separate five-race repeat reproduced all 20 saved observation/action/timed-trace/untimed-settle arrays exactly. The original reference also passed an exact five-race baseline check before calibration. The bounded investigation tested exactly three fixed offsets (+0.0125, +0.025, +0.0375), with all outcomes retained; 55 races and 27,843 control steps included baseline, screens, official sweep and confirmation. No physics changes were made.

These original five starts were used for selection: they are not held-out robustness evidence. Other command-grid settings included falls and a timeout; this is not an all-speed completion claim. Local compatibility and repetition do not guarantee identical server results or future event eligibility.

Rights and publication status

Pollen Robotics’ upstream weights are credited. No standalone open-source license is asserted for them. This Hugging Face publication is an attributed, service-scoped remix under the applicable platform terms, not a blanket off-platform redistribution grant. No new license is invented here.

Published with user authorization; no Arena entry has been submitted.