tags:
- microduck
- microduck-policy
- robotics
- onnx
library_name: microduck
pipeline_tag: robotics
base_model: suraj-phanindra/microduck-sprint-hc5-aifund
base_model_relation: finetune
dfreq
A neural-feedback policy for the Microduck Arena 2 m Sprint, published by djduecaster. This is the V45 physical-leg-adapter version of our 0.856-second local result, not a new speed record.
Local verification
On 25 October 2026, the unchanged artifact was checked with the pinned official Arena CPU evaluator (runner version 4):
- Mean finish time: 0.856 s, 5/5 finishes, selected command 1.15 m/s.
- Individual times: 0.86, 0.86, 0.86, 0.86, 0.84 s.
- Verification covered the complete official 21-command scan and 12 additional finalist races. A separate five-start replay exactly matched the previously saved observations, actions, race traces and settling traces.
- These five starts were used during optimization. This is an in-sample result, not a held-out robustness claim. Most other commanded speeds failed in the scan; this is a specialized sprint policy, not a generally reliable walker.
- Local timing is not a server leaderboard result. The Space evaluates the submitted revision independently, and hardware/software differences can change the outcome.
Controller and provenance
Input: the standard 61-dimensional observation, including actual previous actions; output: 14 joint-position actions, at 50 Hz. Observation normalization is included. Runtime control uses neural feedback and a physical-state leg adapter; it contains no explicit clock counter, seed-ID input, seed-fingerprint selector or trajectory lookup table.
The graph retains the frozen hc5 base, adds our trained residual, and adds a small physical-state adapter affecting knee/ankle outputs. The residual's training lineage used sanitized outputs from game-sprinter-remix4. That teacher uses clock/seed-specific mechanisms; those mechanisms are not present in this deployed graph. Subsequent residual/adapter optimization used native Arena outcomes on the original five starts.
Credit: suraj-phanindra (hc5), juicenv and namenu (base lineage), jrubiosainz, dashi and Muhammadsafi (teacher lineage), and Pollen Robotics (robot, training ecosystem and Arena).
Artifact and limitations
policy.onnx: 1,137,169 bytes
SHA256: 4aa8b27e4e02fbe9dcea1cae2d07f9505b51a62418dad853d2ff55e6a924a243
Simulation only. No physical-robot validation or safety claim. No training code or external runtime helpers are required to load this ONNX; no claim is made that all training code is published here.
The inspected upstream model repositories do not declare standalone model licenses. This repository does not relicense inherited weights as MIT, Apache, or another open-source license. Review the linked upstream terms and Hugging Face Terms of Service, including their public-repository provisions, before reuse. A general off-platform redistribution license is not asserted.