inspect-robots-xpolicylab

Drive any XPolicyLab-served policy from Inspect Robots.

XPolicyLab unifies 40+ VLA and imitation-learning policies (π0 / π0.5, GR00T N1.7, OpenVLA-OFT, RDT-1B, SmolVLA, ACT, Diffusion Policy, …) behind one policy-server contract: each policy runs in its own conda/uv environment and serves inference over a websocket protocol. This plugin is the Inspect Robots side of that wire: a Policy adapter that connects to a running XPolicyLab policy server, so the whole policy zoo becomes evaluable with any Inspect Robots embodiment and task.

The adapter speaks XPolicyLab's msgpack-over-websocket protocol directly — it does not depend on the xpolicylab package (which is not on PyPI). Only the policy server needs an XPolicyLab checkout.

Install

pip install inspect-robots-xpolicylab

The xpolicylab policy then appears in inspect-robots list policies.

Quickstart

Terminal 1 — serve a policy from your XPolicyLab checkout (its own environment, possibly another machine; see the XPolicyLab README for per-policy install/checkpoint details):

cd XPolicyLab/policy/Pi_0
bash setup_eval_policy_server.sh <bench_name> <task_name> <ckpt_name> \
    <env_cfg_type> <action_type> <seed> <policy_gpu_id> <policy_env> 19000 0.0.0.0

Terminal 2 — evaluate it with Inspect Robots:

inspect-robots run --task my-task --policy xpolicylab --embodiment isaacsim \
    -P url=ws://gpu-box:19000 -P cameras=cam_head:base_rgb

or programmatically:

from inspect_robots import eval
from inspect_robots_xpolicylab import XPolicyLabPolicy

with XPolicyLabPolicy(url="ws://gpu-box:19000", cameras="cam_head:base_rgb") as policy:
    log = eval("my-task", policy, "isaacsim")

Constructing the policy and reading .info never touch the network — the websocket connects on first reset()/act(), with retries while the policy server cold-starts (loading VLA weights can take minutes).

Configuration

Mapping-valued args accept a compact string form for the CLI: -P cameras=cam_head:base_rgb,cam_wrist:wrist_rgb.

Default state_map (all XPolicyLab state fields are optional; mapped keys missing from an observation are simply not sent):

Action mapping

Each infer reply is an action chunk — one dict per future control step. Dicts flatten to vectors in a fixed order: for each arm (left then right), the arm key then the end-effector key. The default single-arm joint profile is concat(arm_joint_state, ee_joint_state) → dim 8, matching the inspect-robots-isaacsim default Franka profile.

Lifecycle notes

• reset(scene) ends the previous trial server-side (trial_end) and starts a fresh trial_id; close() ends any open trial and says goodbye.
• eval() closes embodiments it resolves, not policies — use with / close() in programmatic runs. A best-effort atexit hook covers registry-resolved CLI runs.
• If the socket drops mid-eval, the trial in flight errors and the next reset()/act() reconnects (replaying the hello handshake).

Protocol compatibility

Validated against XPolicyLab commit fe71eb54675cef495fea817a637386a4f4529153. The protocol carries no version field; if upstream changes the wire format, _protocol.py is the single small module to diff and update.