tug-pool2
Mjlab-TugOfWar-MicroDuck
A perpetual policy for the microduck (61-D observation, 14 actions, 50 Hz). Runs until told otherwise — a gait, loaded into a policy slot.
Run it on a robot
sudo robotctl policy load <slot> AndoShy/microduck-tug-pool2
The observation normalizer is baked into policy.onnx; feed raw observations.
manifest.json follows schema 2 of the microduck policy manifest (docs/policy-manifest.md in the daemon repo).
Training
- task_id:
Mjlab-TugOfWar-MicroDuck - repo:
git@github.com:pollen-robotics/microduck-challenges - branch:
andoshy-tug - commit:
e8b7d782e - checkpoint:
450 - seed:
1 - base:
mjlab-microduck 0.1.0 @ 1efd853be - started:
2026-10-27T18:02:26Z
Reproduce
Same code, same uv.lock, same command, same seed. Training it again yields a comparable policy, not the same weights: GPU reinforcement learning is not bit-reproducible across machines.
git clone git@github.com:pollen-robotics/microduck-challenges
cd microduck-challenges
git checkout e8b7d782e
uv sync
uv run train Mjlab-TugOfWar-MicroDuck --env.scene.num-envs 2048 --agent.max-iterations 600 --agent.seed 1 --agent.save-interval 50 --agent.logger tensorboard --agent.run-name pool2 --agent.resume True --agent.load-run 2026-10-27_17-10-29_winmark --agent.load-checkpoint model_300.pt --env.actions.opponent.pool logs/pool/ref_t20-500.onnx,logs/pool/ref_t20-500.onnx,logs/pool/ref_t20-500.onnx,logs/pool/leader_nico.onnx,logs/pool/leader_nico.onnx,logs/pool/greecree_v3_mlp.onnx,logs/pool/juicenv.onnx,logs/pool/juicenv.onnx
- Downloads last month
- -