Skip to main content

Configuration

Every run combines a robot, a policy and a backend. Robots and backends can be given by builtin name (--robot g1) or by YAML path; policies are always a path. --print-config shows the fully resolved result.

Robot​

configs/robots/<name>.yaml describes the physical robot and is shared by every policy that runs on it.

KeyMeaning
nameBuiltin name
joint_names, body_namesCanonical joint and body order used across the runtime
modelMJCF path or hf://owner/repo@revision/path.xml
root_bodyFloating base body
initial_q, initial_positionSimulation start state
effort_limit, joint_armature, ...Fixed dynamics and torque limits

Builtin robots: g1, rp1 (28 DoF), rp1-24dof, rp1-parallel, rp1-24dof-parallel. List them with --list-robots.

Policy​

A deploy checkpoint is policy.onnx + policy.yaml:

KeyMeaning
model_pathONNX file; defaults to the YAML's sibling policy.onnx
onnx.in_keys / onnx.out_keysInput/output tensor bindings
joint_names_simulation, body_names_simulationNames the policy was trained with
policy_joint_namesJoints driven by the action, in action order
default_joint_pos, joint_kp, joint_kdPer joint; maps accept regex keys such as .*_knee_joint
action_scale, action_clipAction mapping
compatible_robotsOptional guard against loading on the wrong robot
observationObservation groups; see Observations

Joints are always matched by name, never by index, so a policy trained with a different joint order works unchanged.

Training-only fields​

Exports often carry fields that only matter during training: observation noise (noise_std, bias_noise_std), domain-randomization ranges, curriculum settings, and historical physics overrides (joint_armature, joint_effort_limit, ...). Bolt Runtime accepts these so that original exports load as-is, but never applies them: deployment adds no noise and no randomization, and physics comes from the robot config. Physics overrides print a notice when ignored.

Legacy exports​

Older exports keep their ONNX bindings in a JSON sidecar next to the YAML; this is still read. New exports should embed onnx.in_keys/out_keys in the YAML.

Backend​

--backend mujoco simulates; --backend real drives the hardware described by the robot's hardware section (driver, calibration, rates; see Real robots). A backend YAML path can also be given, e.g. for the replay backend; see Backends.

Paths and assets​

  • Relative paths inside a YAML resolve against that YAML's directory; paths on the command line resolve against the current directory.
  • hf://owner/repo@revision/path resolves through the Hugging Face cache and is downloaded on first use; see Installation.
  • --robot-model overrides the robot's model for a single run.