MicroDuck walking policy (official pretrained copy)
This repository contains BEST_alpha_walking.onnx, an unchanged copy of the
pretrained walking policy distributed by Pollen Robotics. It was downloaded
and smoke-tested locally; it was not trained or fine-tuned by this uploader.
Source
- Original project: pollen-robotics/microduck-simulator
- Pinned revision:
183f99a40bd7308da3e848de961ed32bb02624a5 - Original model file
- SHA-256:
e36332d383997d51401897734cd3e79cf5038406feddb18b4d57ecfb141daa6c
No new license is granted by this copy. Consult the original project and its authors for applicable model usage and redistribution terms.
Interface and local validation
- Input:
obs, float32, shape[1, 61]. - Output:
actions, float32, shape[1, 14]. - Observation normalization is included in the graph.
- ONNX graph checker passed.
- All 512 wide-distribution random input samples produced finite outputs.
- Local CPU inference averaged approximately 0.027 ms per call over 1000 calls. Timing is specific to the test machine.
See validation_report.json for recorded results. No source-checkpoint numerical
parity comparison, physics rollout, or real-hardware test was performed.
Minimal inference example
Install numpy and onnxruntime, download the ONNX file, then run:
import numpy as np
import onnxruntime as ort
session = ort.InferenceSession(
"BEST_alpha_walking.onnx", providers=["CPUExecutionProvider"]
)
obs = np.zeros((1, 61), dtype=np.float32)
actions = session.run(["actions"], {"obs": obs})[0]
print(actions.shape) # (1, 14)
Zero observations here are synthetic smoke-test data. A robot rollout needs the upstream observation layout, control loop, and action processing; these raw outputs are not direct hardware commands.
- Downloads last month
- -