AI GitHub
OpenPI for experimenting with vision-language-action policies
Run and adapt published robot policy examples and checkpoints.
Key Takeaways
- OpenPI connects observations and task descriptions to predicted robot actions.
- Choose a checkpoint and adapter for a documented environment.
- Reproduce a documented simulation or offline example.
- Checkpoints are tied to training distributions and action conventions.
How it works
OpenPI connects observations and task descriptions to predicted robot actions. Robot-specific adapters interpret inputs and outputs for the supported environment.
What you need
Choose a checkpoint and adapter for a documented environment. Validate dimensions and normalization on stored examples before evaluating live control.
What to try first
Reproduce a documented simulation or offline example. Check observation formatting and action normalization with saved data before connecting a robot.
Use a saved example to check which action dimension changes for a small movement. A swapped axis or a sign change should be visible before hardware is involved.
Where to look for failures
Checkpoints are tied to training distributions and action conventions. Validate a new embodiment in a controlled environment and review hardware safety independently.
Related Franklin pages
Start with saved data or a simulator before deploying a policy on hardware. Verify observation layout, action conventions and task success using a known reference episode.
- Train and evaluate manipulation policies in simulated tasks with visual and state observations. ManiSkill for robot manipulation experiments
- Study manipulation tasks in simulated kitchens with varied scenes and objects. RoboCasa for simulated household manipulation
Franklin’s robotics coverage gives examples of where simulation, training and hardware meet. Franklin’s robotics coverage
Project documentation
Try it for your workflow
Explore the project’s documentation for setup and usage, or save this page to a collection for later.
Read the project documentation ↗