Skip to main content
Pass a model’s ID to PolicyClient. Each model runs on its own server.

Comparison

π0.5

Physical Intelligence’s π0.5. It writes the robot state into its language prompt, which helps it follow instructions in scenes it hasn’t seen before. Start with π0.5 for a single- or dual-arm robot with joint-position control.

π0

The earlier Physical Intelligence policy. It has the same observation and action format as π0.5, so you can switch between the two by changing the ID. It’s useful as a baseline or when a fine-tuned π0 adapter already exists for your robot.

GR00T N1.7

NVIDIA’s Isaac GR00T, served with the DROID embodiment: a Franka arm with an exterior camera and a wrist camera. Each call returns end-effector, gripper and joint actions, so you can drive whichever interface your controller accepts. It takes structured observations rather than flat keys. See the GR00T tab on the Observations page.

Fine-tuned adapters

π0 and π0.5 servers can serve LoRA adapters fine-tuned for a specific robot or task, alongside the base model. Select one per request with sampling_params.lora. See Request options. You can only select adapters registered on the server.

Checking a server at runtime

The handshake is the source of truth for a server’s settings. If the docs and the metadata disagree, trust the metadata: