Policy Inference

Run a fine-tuned VLA policy on the robot in closed-loop at a fixed frequency.

1. Install policy dependencies

Policy inference needs openpi and lerobot. They're opt-in (the install is large and conflicts with the docs extra):

uv sync --group pi

2. Pick a station and policy

Env var	Purpose	Where defined
STATION	Robot hardware config (arm, gripper, cameras)	examples/cfg/
POLICY	Policy config (checkpoint, transforms, action space)	examples/policy_cfgs/

Available policies:

POLICY	Model	Notes
Pi05Cfg	openpi pi0.5 (DROID-style)	3 cameras, 32-dim actions, joint_vel
SmolVLACfg	lerobot SmolVLA	2 cameras, joint_pos

List available stations and policies:

uv run -m examples.cfg
uv run -m examples.policy_cfgs

3. Point the policy at your checkpoint

The three fields you'll typically set live at the top of each policy config:

Field	Pi05Cfg	SmolVLACfg
policy_path	Dir with params/ + assets/	lerobot checkpoint dir
asset_id	Subfolder under <policy_path>/assets/ containing norm_stats.json	—
instruction	Language prompt for the task	Language prompt (optional)

You can either edit the defaults in examples/policy_cfgs/pi05.py / smolvla.py, or override them on the CLI (see below).

The action_space in the policy config automatically sets the arm controller — no need to edit station files.

4. Run

# pi0.5
STATION=Xarm7EEFStation POLICY=Pi05Cfg uv run -m examples.policy_inference \
  --policy-path /data/ckpt/pi05_multi_fold \
  --asset-id robotio/multirobot_fold \
  --instruction "Fold the shirt in half."

# SmolVLA
STATION=Xarm7EEFStation POLICY=SmolVLACfg uv run -m examples.policy_inference \
  --policy-path ckpts/smolvla \
  --instruction "Place the cup on the shelf."

Press Enter when prompted to start the loop.

Common flags

Flag	Default	Description
--policy-path	from config	Checkpoint directory
--asset-id (pi0.5)	from config	Norm-stats asset id
--instruction	from config	Language instruction
--freq	50	Control frequency (Hz)
--visualizer	None	Set to Rerun for live visualization
--policy-node-cfg.chunk-size	16	Actions per chunk
--policy-node-cfg.chunk-request-threshold	0.1	Fraction consumed before next chunk request

Inference loop

1. Collect an observation (camera images + proprioception)
2. Send it to the policy server
3. Receive an action chunk and step through it
4. Request a new chunk before the current one is exhausted