# Run Benchmark

> Launch a geniesim_benchmark task locally (typically inside the GUI Docker container) against a user-provided inference server, using the `geniesim benchmark run` CLI verb. Trigger: When the user asks to "run geniesim", "本地跑仿真", "启动仿真任务", "run a benchmark", "launch <some>_<config>.yaml", or wants to execute a benchmark task config (anything under `geniesim_benchmark/config/*.yaml`) against a remote inference host (ip:port).

- Skill: `agibottech/run-benchmark` (Agent Skill)
- Install (CLI): `npx skillmds@latest add agibottech/run-benchmark`
- Raw SKILL.md: https://api.skillmd.com/api/skills/agibottech/run-benchmark/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: DevOps & Infra
- License: MPL-2.0
- Author: agibottech (https://skillmd.com/u/agibottech)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/agibottech/run-benchmark

---


## When to Use

- User wants to run a `geniesim_benchmark` task on their workstation (not the Challenge platform).
- User has an inference server already running somewhere reachable and provides its `ip:port`.
- User references any task config under `source/geniesim_benchmark/src/geniesim_benchmark/config/`.

Do **not** use for:
- Submitting jobs to the Challenge platform → `challenge-submit-job`.
- Verifying an inference server is healthy → use the `check-inference` skill.
- Adding a new benchmark task → `add-benchmark-task`.

## Critical Patterns

1. **Always collect three required inputs first**:
   - Task config (basename, full path, or substring — the CLI resolves all three).
   - Inference IP.
   - Inference port.
2. **The runtime needs `omni_python` / Isaac Sim on the host.** Inside the
   Genie Sim Docker image (`geniesim docker up` → `geniesim docker into`),
   that's already the case. Outside the container the user needs Isaac Sim
   installed system-wide.
3. **Working directory**: anywhere under the repo root works — the CLI
   walks up to find `scripts/` and uses `find_spec` to locate the
   benchmark package.
4. **Confirm before launching.** The task spawns a full simulator and
   typically holds a GPU; ask before kicking it off if there's any ambiguity.

## Workflow

### Step 1 — Collect inputs

If the user hasn't named a config, list candidates:

```bash
geniesim benchmark categories      # show category counts
geniesim benchmark robots          # show robot counts
geniesim benchmark list --robot=<R> --category=<C>
```

Then ask via `AskUserQuestion`:
- **Task config**: free-text (use the basename — e.g. `g2op_if_pick_block_color`).
- **Inference host** as `ip:port`.

### Step 2 — Probe inference (optional but recommended)

Before sinking minutes into Isaac Sim startup, sanity-check the server (uses the
bundled corobot payload):

```bash
geniesim benchmark check-inference --infer-host=<IP>:<PORT>
```

See the `check-inference` skill to override the payload.

### Step 3 — Run the task

Inside the GUI container (`geniesim docker into`):

```bash
geniesim benchmark run <CONFIG> --infer-host=<IP>:<PORT>
```

Example:

```bash
geniesim benchmark run g2op_if_pick_block_color --infer-host=<IP>:<PORT>
```

### Step 4 — Pass-through overrides (when asked)

`geniesim benchmark run` forwards any unknown `--key=value` to the
benchmark's `ParameterServer`. Common ones:

| Flag | Meaning |
|---|---|
| `--app.headless=true` | No GUI (required on remote / batch hosts) |
| `--benchmark.num_episode=N` | Override episode count |
| `--benchmark.seed=N` | RNG / instance-sampling seed |
| `--benchmark.record=true` | Persist episode logs to `output_dir` |
| `--benchmark.policy_class=…` | Use a different policy class |

Full schema: `source/geniesim_benchmark/src/geniesim_benchmark/config/params.py`.

## Commands (copy-paste summary for the user)

```bash
# Host — start the container (GUI by default; add --headless on remote/batch hosts)
cd /path/to/main
geniesim docker up

# Host — drop into a shell inside the running container
geniesim docker into
# inside container:
geniesim status                                       # verify the stack is healthy
geniesim benchmark check-inference --infer-host=<IP>:<PORT>
geniesim benchmark run <CONFIG> --infer-host=<IP>:<PORT>
```

## Notes

- If `geniesim` isn't on `$PATH` (the launcher wasn't installed), substitute `python3 -m geniesim_cli benchmark …` — same args, same behaviour.
- `<CONFIG>` accepts the bare basename (`g2op_if_pick_block_color`), a full path, or a unique substring.
- For batch evaluations, prefer `geniesim benchmark batch --category=… --robot=…` over a shell loop — it forwards extras consistently and prints a per-config pass/fail summary.
- The new CLI replaces the older ad-hoc `omni_python app/app.py --config …` invocation. The new form normalizes interpreter selection, host shorthand, and config resolution.