gemini-robotics git:20261104.1687817A

Immutable. This exact content never changes and is served at /api/v1/blob/76da696ae364cb09.

---
name: gemini-robotics
description: Use when working on Gemini Robotics 2 toolRefs — ER embodied-reasoning planning or rubric evaluation via the hosted Gemini API.
---

# Gemini Robotics

Gemini Robotics is the closed-weight VLA tool for ER (embodied reasoning)
planning and rubric evaluation, all served through the
hosted Gemini API (``GOOGLE_API_KEY`` required).

The toolRefs use endpoint and model defaults live-validated 2026-10-10.
Workflow specs expose ``config.api_base_url`` and ``config.model_id`` as
optional overrides; omitted or empty values defer to the runtime defaults.

Workflow tasks use the default CPU image with staged NPA source because the
adapter calls a hosted API. Pass `--secret-env GOOGLE_API_KEY` when submitting;
the renderer declares that credential hint and does not select a local model
image. The historical validation does not prove current provider availability;
transport tests remain local contract evidence only.

## Interfaces

- CLI: `npa workbench gemini-robotics <plan|eval|distill> --help`
- `distill`: teacher-student loop - ER-2 plans, MuJoCo generates demos,
  PyTorch behavior-clones a student policy (CUDA when available; run on a
  Nebius GPU job), MuJoCo evaluates closed-loop. One command, no stubs.
- Python SDK (workbench-first): `npa.workbench.gemini_robotics`
  (`plan`, `eval` — aliases of `run_er_planning_stage`,
  `run_eval_stage`)
- Workflow module: `npa.workflows.byof.gemini_robotics_pipeline`
  (stage functions, receipt writing, argparse entrypoint)

## Conventions

- The CLI lives at `npa.cli.workbench.gemini_robotics` (multi-command package
  form for new tools); the SDK surface lives at `npa.workbench.gemini_robotics`
  and does not go through `make_cli_wrapper`.
- Public workflow inputs and outputs are exact `s3://` objects. `plan` accepts
  `config.task` and writes `config.output_uri`; `eval` accepts an `input_uri`
  JSON object with non-empty `plan_text` and `rubric` strings and writes
  `config.output_uri`. Eval receipts record the exact input URI, ETag, and
  SHA-256; all receipts use conditional create and are not overwritten.
- Plans and evaluations are advisory only. The adapter does not execute a
  safety-review tool and does not authorize physical execution.