Skip to main content

Listed

, home
FREN

Policy

In French:
Politique (policy)

Definition

A function that maps what the robot perceives to the action it should take. It can be a simple table or a large neural network that receives images and an instruction in natural language. Training a robot by imitation or by reinforcement amounts to learning its policy.

Checked on against the references below. · Written with AI · checked against the sources Source: Sutton and Barto, Reinforcement Learning: An Introduction, § 1.3 Elements of Reinforcement Learning.

Concrete examples

Hello Robot describes a repository for the Stretch 4 that uses LeRobot to record demonstrations on the robot, train a policy on a desktop GPU, then run it on the robot. LimX Dynamics publishes a repository for the TRON 2 devoted to serving the pi0.5 policy and fine-tuning it on a task, with a server profile that groups the model checkpoint, the prompt and the port.

What it changes for you

On the Stretch 4, not every policy deploys in the same place: Diffusion and ACT are tested and can run on the optional onboard Jetson, while pi0.5 fine-tuned with LoRA is tested but has to run on a desktop GPU. Other LeRobot policies are listed there as untested.

Additional material checked on .

Related terms

References

Back to the glossary