almuli/microduck-sprint-race4

sprint-race4

2 m Sprint race env (settle, line, Arena fall rule) fine-tuned from hc5 (suraj-phanindra, juicenv, namenu); no clock or seed tricks

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> almuli/microduck-sprint-race4

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).

A remix of suraj-phanindra/microduck-sprint-hc5-aifund.

Training

• task_id: Mjlab-Sprint2m-MicroDuck
• repo: https://github.com/pollen-robotics/microduck-challenges
• branch: sprint-race
• commit: d810a5bbd
• checkpoint: 550
• seed: 1
• base: mjlab-microduck 0.1.0 @ 1efd853be
• started: 2026-10-28T15:57:17Z

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
cd microduck-challenges
git checkout d810a5bbd
uv sync
uv run train Mjlab-Sprint2m-MicroDuck --env.scene.num-envs 4096 --agent.max-iterations 800 --agent.seed 1 --agent.logger tensorboard --agent.run-name race4 --agent.resume True --agent.load-run finetune_r3_250 --agent.load-checkpoint model_0.pt --agent.save-interval 50 --agent.algorithm.learning-rate 1e-4 --agent.algorithm.desired-kl 0.005 --agent.algorithm.entropy-coef 0.002