YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Microduck autonomous jump policy v24
Simulation-only policy and a fixed deployable action filter trained in
pollen-robotics/microduck_rl.
The actor maps the shared 61D onboard observation contract to 14 relative joint targets at 50 Hz. Runtime observations are IMU angular velocity, projected gravity, joint positions/velocities, previous actions, and a one-shot phase clock. The actor receives no MuJoCo root pose, foot-contact state, target coordinates, reference motion, or external motion commands.
deployment.json describes the 11 fixed phase-gated gains and biases applied
to actor output. They were optimized in simulation, are independent of runtime
simulator state, and require only the same phase clock already present in the
actor observation.
Fresh randomized qualification
- Episodes: 4,096 standing-only starts
- Strict success: 7.568% (310/4,096)
- Controlled success: 5.835% (239/4,096)
- Valid two-foot flight: 60.986%
- Falls: 41.504%
Strict success requires at least 4 cm trunk rise, 100 ms with both feet
airborne, and then 300 ms of continuous two-foot stability within the height,
tilt, drift, linear-speed, and angular-speed thresholds recorded in
evals/evaluation.json.
This is a real improvement and can produce clean accepted runs, but a 7.57% success rate is not hardware-ready. No hardware qualification has been run.
Files
checkpoints/microduck-jump-v24-policy.onnx: 61D → 14D inference actor.checkpoints/jump-elite-distilled-r2.pt: source PyTorch checkpoint.controller/deployment.json: runtime action-filter contract.controller/jump_action_filter_v24.json: complete CEM optimization record.evals/evaluation.json: aggregate qualification and thresholds.evals/episodes.csv: all 4,096 per-episode results.videos/: accepted-run edits, telemetry, and evidence sidecars.