inspect-robots 0.6.0

An evaluation framework for VLA (vision-language-action) models across real robots and simulators — the Inspect AI for robotics.

Inspect Robots

An open-source evaluation framework for physical AI and VLA (vision-language-action) models

Define a robotics benchmark once, then run any policy against any compatible embodiment (a real robot or a simulator) with reproducible logs and first-class Rerun visualization.

If you know Inspect AI, this is that for robotics.

Documentation · Quickstart · Concepts · For LLMs

Note: This project is in early development. The API may change between releases, so pin a version before depending on it.

---

One framework, two swappable inputs

LLM evaluations have a single swappable input: the model. Robotics evaluations have two, and Inspect Robots makes both first-class and orthogonal:

Policy: the VLA	The "brain". Maps an observation + instruction to an action chunk (a horizon of actions executed open-loop, as π0 / ACT / diffusion policies do).
Embodiment: the robot or sim	The "body + world". Produces observations, executes actions, owns the action/observation spaces and control rate. Real-robot-first; sims are a stricter special case.

A Task, a dataset of Scenes (initial conditions, instructions, success targets) plus scorers, is defined independently of both. Before any rollout, Inspect Robots checks the (policy, embodiment) pair is compatible (action/observation spaces, semantics, control rate, scene realizability) and fails fast if not.

Install

In a fresh directory (or your existing project), create a virtual environment and install (system Pythons on modern distros reject bare pip install, per PEP 668):

uv venv && uv pip install "inspect-robots[rerun]"

The rerun extra powers the live run viewer. For the numpy-only core:

uv venv && uv pip install inspect-robots

Any venv workflow works the same way (python3 -m venv .venv and that venv's pip and console scripts); with uv, run commands through uv run as shown below and no activation is needed.

Quickstart

Set your defaults once. The policy and embodiment come from installed plugins (inspect-robots-yam shown here); replace the three camera paths with your rig's V4L2 color nodes:

mkdir -p ~/.config/inspect-robots && cat > ~/.config/inspect-robots/config.ini <<'EOF'
[defaults]
policy = molmoact2        # from the inspect-robots-yam plugin
embodiment = yam_arms     # same plugin; cameras configured below
scorer = success_at_end
max_steps = 1200          # 120 s at 10 Hz
rerun = true              # live viewer of cameras/state/actions each run
store_frames = true       # save each run's camera frames under logs/frames/

[embodiment.args]
top_cam_device = /dev/v4l/by-id/YOUR-TOP-CAM
left_cam_device = /dev/v4l/by-id/YOUR-LEFT-CAM
right_cam_device = /dev/v4l/by-id/YOUR-RIGHT-CAM
EOF

Then tell the robot what to do:

uv run inspect-robots "place the fork on the plate"

Every run opens a live Rerun viewer streaming the cameras, proprioception, and actions straight from the eval pipeline, so you watch exactly what the policy sees while the robot moves. CLI flags override any default (--no-rerun, --no-store-frames, --max-steps 300, ...), and the same instruction runs on your configured simulator instead of the real robot:

uv run inspect-robots "place the fork on the plate" --sim

The full command line resolves any registered task/policy/embodiment (builtins + installed plugins). List what is registered:

uv run inspect-robots list

Run a registered task with explicit components:

uv run inspect-robots run --task cubepick-reach --policy scripted --embodiment cubepick

Pretty-print a saved eval log:

uv run inspect-robots inspect logs/cubepick-reach_*.json

And everything is a Python API. No hardware or simulator needed: the dependency-free CubePick mock world exercises the whole stack:

from inspect_robots import eval
from inspect_robots.mock import CubePickEmbodiment, ScriptedPolicy
from inspect_robots.scene import Scene
from inspect_robots.scorer import success_at_end
from inspect_robots.task import Task

task = Task(
    name="cubepick-reach",
    scenes=[Scene(id=f"layout-{i}", instruction="reach the cube", init_seed=i) for i in range(5)],
    scorer=success_at_end(),
    max_steps=80,
)

# The two swappable inputs: a policy (VLA) and an embodiment (robot/sim).
(log,) = eval(task, ScriptedPolicy(), CubePickEmbodiment())
print(log.status, log.results.metrics)   # success {'success_at_end': 1.0}

Why Inspect Robots

• Real-world first. Interfaces assume real-robot reality: human-in-the-loop reset, no privileged success oracle, wall-clock control rate. Simulators just offer more (seeding, privileged success, rendering) via opt-in capabilities.
• Reproducible. Every run yields an immutable, schema-versioned EvalLog with the resolved config, git revision, and package versions. It is re-readable across releases and re-scorable offline.
• Light core. Depends only on NumPy. Rerun and simulator/VLA backends are optional extras and separately installable plugins.
• Safe unattended. An explicit error taxonomy separates "record and continue" from "halt and require a human", so a faulted robot never auto-advances overnight.
• Rerun visualization. Stream camera images, 3D poses, joint/action time-series, and success markers to a .rrd recording.
• Pluggable. Ship inspect-robots-maniskill or inspect-robots-openvla as separate packages. Entry points make them appear in inspect-robots list automatically.
• VLA-native. Action chunking, open-loop execution, and ACT/ALOHA temporal ensembling are built in, with action semantics (control mode, rotation representation, gripper, frame) that make compatibility and ensembling correct.

First-party plugins

Both halves of an eval (the "body" and the "brain") have a ready-made adapter shipped from this repo as separate packages:

• inspect-robots-isaacsim: run evals against an Isaac Lab simulation (--embodiment isaacsim).
• inspect-robots-xpolicylab: drive any XPolicyLab-served policy. One adapter puts its zoo of 40+ VLAs (π0/π0.5, GR00T, OpenVLA-OFT, RDT-1B, SmolVLA, ACT, …) behind --policy xpolicylab -P url=ws://gpu-box:19000.
• inspect-robots-agent: let a frontier LLM (Claude, GPT, anything behind an OpenAI-compatible API) drive any embodiment through tool calls, as a first-class policy. The same --policy agent runs ad-hoc instructions and scores on registered tasks next to fine-tuned VLAs.

# Isaac Lab world + a π0 checkpoint served by XPolicyLab, evaluated end to end:
inspect-robots run --task my-task --embodiment isaacsim \
    --policy xpolicylab -P url=ws://gpu-box:19000 -P cameras=cam_head:base_rgb

# Claude driving the mock world, no hardware or GPU required:
export ANTHROPIC_API_KEY=sk-ant-...
inspect-robots "pick up the cube" --policy agent \
    -P model=anthropic/claude-fable-5 --embodiment cubepick

Safety guardrails (a bounds clamp plus a per-step delta limit derived from the embodiment's action space) are wired into every CLI run by default, for every policy. Turning them off requires an explicit --disable-guardrails. Persist your usual setup once with inspect-robots config set embodiment NAME and inspect-robots config set policy NAME, then a bare inspect-robots "wipe the table" does the rest.

How it maps to Inspect AI

If you know Inspect AI, you already know Inspect Robots.

Inspect AI	Inspect Robots
Model	Policy (VLA) + Embodiment (two inputs)
Task = dataset + solver + scorer	Task = scenes + controller + scorer
Sample	Scene
Solver chain	Controller middleware (chunking, ensembling, smoothing)
eval() → EvalLog	eval() → EvalLog
@task / @solver / @scorer + registry	@task / @policy / @embodiment / @scorer + entry points

This repository is the framework. Concrete benchmarks live in WorldEvals, the benchmark catalog, and backend adapters live in separate plugin packages.

Documentation

Full guides and an auto-generated API reference live at inspectrobots.org. LLM-friendly versions: llms.txt and llms-full.txt.

Development

Dependency changes: after editing dependencies in pyproject.toml, run uv lock and commit the updated lockfile. CI installs with uv sync --locked and fails with "the lockfile needs to be updated" if you forget. Day-to-day conventions (PR-only main, the required ci-ok check, one-click releases) are documented in CLAUDE.md.

uv venv && uv pip install -e ".[dev]"
uv run pre-commit install          # ruff + mypy on commit, 100% coverage on push
uv run pytest --cov                 # 100% coverage required
uv run ruff check . && uv run mypy

Pre-commit hooks and a blocking CI coverage gate keep main green. See CONTRIBUTING.md and the design docs in plans/.

Citation

If you use Inspect Robots in your research, please cite it:

@software{inspect-robots,
  author  = {Robocurve},
  title   = {Inspect Robots: The open-source evaluation framework for physical AI},
  year    = {2026},
  url     = {https://github.com/robocurve/inspect-robots},
  version = {0.3.0},
  license = {MIT}
}

License

MIT