Policy
Status: hardware-verified
Hardware-verified Policy inference and interventions (policy, policy_teleop) are checked on the real Rakuda2. Asynchronous inference, replaying a recording and a policy in another process are Mock only.
A type: policy controller turns a learned policy's actions into follower targets. The policy is an object with reset() and infer(observation): it gets the same observation as the recording and returns absolute joint targets. The controller takes care of the rate, interpolation, speed limits and safety checks.
# nodes/policy.yaml (excerpt)
type: policy
followers: [follower]
factory: roboloom.policy:hold # "module:callable"
observe: [camera]
rate_hz: 15
action_horizon: 8
uv run roboloom infer --seconds 20 --policy my_lab.policies:load
With teleop:, the leader can drive the follower before and during an episode (Teleoperation and interventions).
Supported robots
| Robot | Options | Status | Example |
|---|---|---|---|
| Rakuda2 | policy, policy_teleop |
Hardware-verified | Rakuda2 controllers |
| PiPER + Pika Gripper | policy |
In development | PiPER controllers |
| Your own robot | depends on the configuration | Mock only | Adding your own robot |
In detail
How to write a policy, LeRobot policies, rates and interpolation, and evaluation are in the Policies and inference chapter. The keys are listed in Policy configuration.