Worker protocol

A worker is a robot, simulated or real, behind HTTP: harness serve --factory module:create_environment [--options FILE] [--port P] [--contract C --embodiment E] (src/harness_policy/core/worker.py). One request at a time; loopback by default (binding another interface needs HARNESS_WORKER_TOKEN, sent as a bearer token). src/harness_policy/core/robot.py (RemoteEnvironment) is the client.

POST /rpc   {"method": "...", "arguments": {...}}   ->   {"result": ...} | {"error": "..."}

Method	Arguments	Result
describe		{protocol_version: 1, name, tools}
reset	task, seed	the first observation: {episode_id, step, instruction, state, images}
call	name, arguments	the result of one of the environment’s methods, below

The robot API (call)

Name	Arguments	Result
profile		static facts: joint groups, action limits, control rate, cameras, arm bases, gripper constants, planner limits, shield options, return_to_start, the step limit (native contract)
state	episode_id	step, ended, joint_position (14 on the dual ARX X5: left 6 joints, left gripper, right 6 joints, right gripper), left_ee_pose, right_ee_pose (link6, [x, y, z, qw, qx, qy, qz], env frame)
tick	episode_id, step, command (absolute joint targets), optional cameras	{executed, shield: {verdict, command, reasons}, state[, observation]}
observe	episode_id, cameras, image_format, quality	{state, images (data URLs), intrinsics, camera_to_world (OpenCV convention), image_size}
fk	command (joints)	{arm: link6 pose}
ik	episode_id, arm, pose	{success, joints} from the current joints
plan	episode_id, arm, pose or joints, raw	{success, waypoints} at the control rate (raw: the planner’s own path)
policy_obs	episode_id	the observation RoboDojo’s evaluation client hands a policy (for run_policy)
claim_outcome	episode_id	a token, claimed once by the driver after reset
outcome	token	{success, reward, process_score, terminated, truncated, steps}
action_spec		RoboDojo only: per arm its mount pose, joint limits and velocity limits; control rates

Rules every worker keeps:

• Episode and step. tick names the episode and the step it was computed for; a stale one is rejected and never executed again. A possibly executed motion is never retried silently.
• The shield runs here. Every tick command is checked before it reaches the robot (The episode runtime); programs cannot bypass it.
• The outcome is guarded. Only the holder of the token (the driver) reads it; programs and the model never see it.

Contracts

--contract xpolicylab --embodiment arx_x5 filters every result to what XPolicyLab’s official client hands a policy: grippers as their last commands, intrinsics and camera_to_world from the embodiment, no step limit in the profile or the state (src/harness_policy/core/contract.py). The profile then says observation_contract: "xpolicylab", gripper_feedback: "commanded".

Implementations

Factory	Bench
benches.robodojo.runtime:create_environment	RoboDojo, one task per process (Benches)
benches.robotwin.runtime:create_environment	RoboTwin 2.0, one task per process
harness_policy.connectors.xpolicylab.Bridge	not served by harness serve: the XPolicyLab connector, a worker whose robot is XPolicyLab’s client