Code as Policy

In code as policy, a program is the controller, often written by a code model and revised from execution feedback. The program calls a fixed API of perception queries and robot skills, such as finding a gear, grasping at a pose, or planning a motion, and runs in closed loop against the environment. Because the policy is a program, you can read it, compare it with another program, and run it again. A skill can itself be a motion planner or a server–client VLA, so a program can coordinate the other kinds of policy.

VABAR evaluates this kind of policy in Arena on gear, tool, syringe, and cable tasks from industrial, medical, and infrastructure settings. Its programs are Graph-as-Policy (GaP) graphs, whose nodes call perception models and skills from open-robot-skills.

How the program talks to Arena

In VABAR, the program runs in its own process, next to the simulation, and exchanges one message with Arena per control step:

Arena                                          program process
─────                                          ───────────────
policy ── camera images, joint state ──────►   graph runtime + skills
       ◄─ arm and gripper targets ──────────   (perception, IK, motion planning)

On each step, the policy sends the camera images and the robot’s joint state, and the program replies with joint targets for the arm and gripper. The program owns perception, kinematics, and motion planning. Arena owns the simulation, recording, and scoring, so a program that finishes counts as a success only when the task’s success check passes. With several environments, VABAR starts one program process for each.

To the runners, VABAR’s GapRemotePolicy (see gap_remote_policy.py) is still a PolicyBase. get_action(env, obs) exchanges one step with the program, and reset(env_ids) restarts the program for those environments. The policy runner, the Experiment Runner, metrics, and OSMO workflows need no changes, so code policies are evaluated on the same tasks and variations as VLAs and agents.

Next steps

Return to the policy overview to compare the four kinds of policy. See Arena Experiments for how evaluations run locally and on OSMO.