hermes-death-metal-headbang
Got a Talent performer: Pollen's recipe as is (task Mjlab-GotATalent-MicroDuck, seed 1, 3000 iterations, 4096 envs, one RTX 5090, about an hour). The performance (timeline.json) was written by Hermes Agent from one plain sentence, see the replay. Trained in simulation only. Not yet tested on a real 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> witcheer/microduck-got-a-talent
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-GotATalent-MicroDuck - repo:
https://github.com/pollen-robotics/microduck-challenges.git - branch:
HEAD - commit:
4b675ebf4 - checkpoint:
2000 - seed:
1 - base:
mjlab-microduck 0.1.0 @ 1efd853be - started:
2026-10-28T01:27:23Z
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 https://github.com/pollen-robotics/microduck-challenges.git
cd microduck-challenges
git checkout 4b675ebf4
uv sync
uv run train Mjlab-GotATalent-MicroDuck --env.scene.num-envs 4096 --agent.max-iterations 3001 --agent.save-interval 250 --agent.seed 1 --agent.logger tensorboard --agent.run-name talent
- Downloads last month
- 9