Run Experiment

Run Experiment

Execute an existing Experiment Definition locally and finish with evidence from its canonical artifacts. Keep runtime readiness, regression testing, direct Policy Runner workflows, and managed submission outside this skill.

Read the checked-out documentation

Before composing a command, read:

• docs/pages/concepts/concept_arena_experiments.rst for Experiment structure and override rules.
• docs/pages/quickstart/arena_experiment.rst for the maintained local example.
• docs/pages/quickstart/environment_variations.rst when listing or applying variations.

Treat the current checkout as the source of truth. If this skill differs from the documentation or runner CLI, follow the checkout and report the mismatch.

Use typed YAML as the primary interface. Accept legacy JSON only as a pass-through compatibility path; do not create new legacy configurations or apply Hydra overrides to them.

Preflight the run

1. Confirm the repository root with git rev-parse --show-toplevel.
2. Resolve an explicit Experiment path. Never rely on the runner's legacy default configuration.
3. Inspect the Experiment's Runs, policies, rollout limits, rebuild counts, environment counts, cameras, variations, and referenced files. Summarize unexpectedly large work before starting it.
4. Preserve a native or Docker route already selected by the user. If both are ready and the request does not choose one, use the preferred route in the current installation documentation and state the choice. If neither route is ready, use setup-arena; do not install dependencies, build an image, create mounts, or recreate a container here.
5. Confirm that referenced configs, datasets, checkpoints, and output locations are available from the selected runtime. Identify a policy from each Run's resolved policy.type, not the Experiment filename.
6. For an OpenPI Pi0RemotePolicy or registered pi0_remote policy whose remote_host is local, use serve-openpi-policy for each distinct variant and port. It must reuse a compatible ready server or start one before this workflow launches the Experiment; the user does not need to invoke that skill separately. If Runs request different variants on the same port, stop and ask the user to resolve the conflict.
7. For a GR00T Gr00tRemoteClosedloopPolicy or registered gr00t_remote_closedloop policy whose remote_host is local, use serve-gr00t-policy. It may automatically start the maintained N1.6-DROID server or require an explicit model and embodiment for another client configuration; the Experiment does not declare the server checkpoint. If multiple Runs require incompatible server contracts on one port, resolve the conflict before launching.
8. For a supported policy pointing to another host, or another remote policy without a matching local server skill, require an already-running and reachable endpoint. Do not replace a deliberately remote endpoint with a local server or submit it to managed compute.

For a built-in smoke evaluation, use isaaclab_arena_environments/experiment_configs/getting_started_experiment.yaml. It uses the local zero-action policy and requires no model or policy server.

Select execution options

• Default to --viz none for unattended execution. Use --viz kit only when requested.
• Record viewport or camera video only when requested. Camera recording enables camera support and can materially increase GPU memory and output size.
• Preserve the runner's default stop-on-first-error behavior. Add --continue_on_error only when the user wants the remaining Runs attempted after a failure.
• Use --serve_evaluation_report only when explicitly requested; it binds an HTTP server and keeps the process running until interrupted.
• Use --list_variations as an inspection operation. It does not run rollouts or create Experiment result artifacts.
• Let the runner create a timestamped directory below outputs/ unless the user chooses a base or exact output location. An exact --experiment_output_directory must be missing or empty. Never clear or reuse a nonempty directory. If a requested exact directory is nonempty, stop and offer either a different exact directory or a fresh timestamped child through --output_base_dir; do not silently change the output semantics.

Build the local command

Always pass --experiment_config explicitly. Keep runner flags separate from trailing Hydra overrides.

For native uv, run from the repository root and accept the Isaac Sim EULA non-interactively:

env OMNI_KIT_ACCEPT_EULA=YES ACCEPT_EULA=Y \
  .venv/bin/python isaaclab_arena/evaluation/experiment_runner.py \
  --experiment_config <experiment.yaml> \
  --viz none \
  <override>...

For Docker, first use docker ps to select the single running container that mounts the absolute repository root. Resolve the host username separately with id -un, then substitute both literal values below rather than hardcoding a container name:

docker exec <container> su <host-user> -c \
  "cd /workspaces/isaaclab_arena && \
   /isaac-sim/python.sh isaaclab_arena/evaluation/experiment_runner.py \
   --experiment_config <experiment.yaml> \
   --viz none \
   <override>..."

Use the local override namespace:

shared.rollout_limit.num_episodes=4
runs.parallel_envs.environment_builder.num_envs=8

Do not prefix local overrides with experiment_cfg.; that prefix belongs to OSMO submission. Quote override tokens that contain shell-sensitive characters. Do not edit the source YAML when a declared field can be changed with an override.

Execute and monitor

Run the command in the selected ready runtime and wait for it to finish. Runs execute in YAML order inside one SimulationApp; each Run builds a fresh environment. Report meaningful progress during a long evaluation.

On failure, preserve the logs and partial output. Do not delete artifacts or silently rerun with different settings. By default, an early failure can prevent creation of the canonical result and report. With --continue_on_error, the process can exit zero even when one or more Runs failed.

Verify the outcome

For an evaluation that reaches finalization:

1. Locate the exact Experiment output directory printed by the runner.
2. Require arena_experiment_result.json and index.html.
3. Read every Run's status from arena_experiment_result.json and report completed and failed Runs.
4. Confirm expected episode_results_rebuild<N>.jsonl files under each completed Run directory.
5. Report requested videos and the HTML report path when present.

Treat the canonical result statuses, not process exit alone, as the workflow outcome. Keep policy task success separate from execution success: zero-action episodes are expected to be semantically unsuccessful while still proving that the evaluation pipeline works. Metrics are printed to the console; the runner does not currently promise a separate metrics.json.

For --list_variations, report the catalogue and explicitly state that no rollout or output artifacts were expected.

Finish with the selected runtime, any policy-server endpoint and whether it was reused or started, the Experiment path, effective overrides, output path, per-Run statuses, episode counts, and any preserved partial artifacts or failures.

Hand off other workflows

• Use setup-arena for installation, container creation, mounts, or readiness repair.
• Use run-tests for pytest and regression checks.
• Use the Environment Runner for interactive inspection without a policy.
• Use the Policy Runner for torchrun, external environments, or an explicitly requested direct rollout outside an Experiment Definition.
• Use a separate OSMO submission skill for previews, cluster resources, submission, monitoring, or remote result download. An unqualified request to "run" an Experiment means local execution.
• Use serve-openpi-policy automatically when a local OpenPI Experiment needs a compatible server. Leave that server running after the Experiment unless the user asks to stop it.
• Use serve-gr00t-policy automatically when a local GR00T Experiment needs a compatible server. Leave that server running after the Experiment unless the user asks to stop it.

References

• Evaluations
• OpenPI server skill
• GR00T server skill
• Arena Experiments
• First Arena Experiment
• Environment variations
• Experiment Runner CLI