Microduck headstand: split
Simulation policy for Mjlab-HeadstandKickup-Flat-MicroDuck. The original checkpoint and an ONNX export with the observation normalizer included are provided. Input: 61 observation values; output: 14 joint actions.
The original checkpoint passed 32 of 32 individual simulated attempts at seed 0 under the documented evaluator. This does not establish hardware performance or success from arbitrary starting poses. The ONNX export is provided separately; the published rollout counts were measured with the original checkpoint.
Files:
model_1499.pt: the unchanged evaluated checkpoint.policy.onnx: export through the training repository's standard exporter.provenance.json: source run, hashes, and code provenance.evaluation.json: individual evaluation report.
The six policies together completed 87 of 96 routine attempts across seeds 0, 1, and 2. See the evaluation code, setup, and raw results. The saved checkpoint can be placed in the evaluator's cache at logs/rsl_rl/microduck_headstand_kickup/wandb_checkpoints/y2fllvgj/model_1499.pt; no W&B download is then needed for that checkpoint.
These are simulation research artifacts, not a robot-daemon installation package. No runtime manifest or hardware installation command is supplied. The routine uses contact-based handovers and a separate standing policy, and has not been tested on a physical robot. Head-force samples are taken at 50 Hz and may miss shorter impacts. Exit success checks final standing, not leg shape throughout the exit.