microduck-lx-seed2-cp2999
robot control policy
1 min readupdated
A new standing gait policy for the microduck robot, trained with PPO (Proximal Policy Optimization, a reinforcement learning algorithm). The model takes 61-dimensional observations and outputs 14 control actions at 50 Hz. It is optimized for lane speed, one-sided path tracking, and smooth locomotion, making it a solid fit for robotics developers and researchers working with the Pollen Robotics platform.
Key features
- Trained using the recipe from the public heyskylark experiment log.
- Bundles an observation normalizer directly inside the policy.onnx file.
- Uses the standard version 2 manifest schema for the microduck daemon.
- Reproducible training pipeline via pinned source code and seed.
What you can do
- Deploy the pre-trained policy onto a physical robot using the robotctl utility.
- Retrain a similar policy from scratch via the microduck-challenges repository.
Example
sudo robotctl policy load prithviraj-maurya/microduck-lx-seed2-cp2999Read more
DetailsTask: Mjlab-Sprint2m-MicroDuck
Architecture: 61 inputs, 14 outputs, 50 Hz
Checkpoint: 2999, seed: 2










Comments