inspect-robots-isaacsim

An Isaac Lab (Isaac Sim) embodiment plugin for Inspect Robots — the "Inspect AI for robotics".

Inspect Robots factors a robotics eval into two swappable inputs: a Policy (the VLA "brain") and an Embodiment (the "body + world"). This package supplies the second one, backed by a real Isaac Lab physics simulation, so you can run any compatible VLA against your Isaac Sim setup.

The default profile is a 7-DoF Franka Panda under joint-position control with a binary gripper (action dimension = num_arm_joints + 1 = 8).

Install

This package installs and registers on any machine, but reset()/step() need a working Isaac Lab environment (NVIDIA Omniverse + GPU) — Isaac Sim is not a PyPI dependency and is imported lazily.

# inside the conda/venv that already has isaaclab + isaacsim available
pip install inspect-robots inspect-robots-isaacsim

Use it

The embodiment is discovered through the inspect_robots.embodiments entry point, so it appears in the CLI without any import:

inspect-robots list embodiments          # -> includes "isaacsim"

inspect-robots run \
  --task my-benchmark \
  --policy my-vla \
  --embodiment isaacsim \
  -E task_id=Isaac-Lift-Cube-Franka-v0 \
  -E headless=true

Or programmatically:

from inspect_robots import eval

logs = eval("my-benchmark", "my-vla", "isaacsim")
print(logs[0].results.metrics)

Constructing directly (e.g. for a non-Franka arm or extra cameras):

from inspect_robots_isaacsim import IsaacSimEmbodiment

emb = IsaacSimEmbodiment(
    task_id="Isaac-Open-Drawer-Franka-v0",
    num_arm_joints=7,
    cameras=[("base_rgb", 224, 224), ("wrist_rgb", 224, 224)],
    control_hz=30.0,
    headless=True,
)

Compatibility

Inspect Robots fail-fast-checks the (policy, embodiment) pair before any rollout. To run against this embodiment your policy must emit 8-D joint_pos actions (control_mode="joint_pos", gripper="binary") and require only observation keys this task provides. The mock cubepick policies (2-D eef_delta_pos) are intentionally incompatible — bring a Franka-trained VLA.

Mapping your task

Isaac Lab tasks vary in their observation-dict layout and success signal. The constructor exposes hooks so you don't edit the adapter:

Memory & GPU hygiene

Long unattended evals should not creep in RAM or VRAM. This adapter is built to hold nothing per step, but a few usage rules keep a full run leak-free:

• Free the simulator when done. eval() closes what it resolves: an embodiment looked up by registry name (e.g. embodiment="isaacsim") is closed when the run finishes. An embodiment object you construct is yours to close — an open SimulationApp holds GPU memory until the process exits, so use it as a context manager (or close() in a finally):

  with IsaacSimEmbodiment() as emb:
      eval("my-bench", "my-vla", emb)
  # GPU + sim torn down here
  

  close() is idempotent and safe to call before launch or twice.

• One simulator per process. Isaac Sim is a hard process singleton; if you construct several embodiments the adapter reuses the one live SimulationApp rather than launching (and leaking) a second.

• Stream frames to disk for long episodes. Inspect Robots keeps each step's observation — including camera frames — in the per-trial record. Pass store_frames=True to eval() so frames go to disk side-cars instead of RAM. (Per-trial records are released after each scene is scored, so there is no cross-trial accumulation regardless.)

• Skip rendering you don't need. If your scorer is state/oracle based, build the embodiment with cameras=() to avoid allocating images at all.

The adapter never stores tensors on self, and copies observations off Isaac's reused buffers (.astype), so a step loop holds no growing references — a tracemalloc regression test asserts flat RAM over thousands of steps.

Develop / test

The test suite runs without Isaac (it checks spaces, semantics, protocol conformance, compatibility, and registry wiring):

pip install -e ".[dev]"
pytest -q

License

MIT.