Google DeepMind unveiled Gemini Robotics 2 on July 30, 2026 — a family of three models that extend robot control from a single arm to a humanoid's whole body and to coordinating several machines at once. It is the first generation in which a DeepMind model plans a task, tracks its progress from video, and splits work across different robots.
Key takeaways
- Three models: Gemini Robotics 2 (VLA), Gemini Robotics ER 2 (planning and reasoning), Gemini Robotics On-Device 2 (local variant).
- Whole-body intelligence: control of walking, crouching and manipulation across the humanoid's full kinematic chain.
- ER 2 tracks tasks live from video: 91.3% moment-finding accuracy and sub-second latency.
- On-Device 2 adapts to a new robot in a few hours, typically on fewer than 200 examples.
- Test robots: Apptronik Apollo 2 and Franka Duo. ER 2 is available in Google AI Studio; VLA models are early-access only.
Three models, one division of labor
The family splits into three roles. Gemini Robotics 2 is a vision-language-action (VLA) model — it turns an instruction and an image into motion for a full humanoid or bi-arm robot. Gemini Robotics ER 2 is a reasoning model (VLM?VLM: A model that combines image and language understanding (a vision-language model).) that acts as a high-level brain: it plans multi-step tasks, reasons about space, and distributes work across machines. Gemini Robotics On-Device 2 is a trimmed VLA that runs locally on the robot, without a cloud connection.
What DeepMind calls whole-body intelligence is control of the robot's entire body, not just its gripper. The model coordinates locomotion — walking, crouching, reaching — with object manipulation across the full kinematic chain. In earlier generations the model mostly drove the arm and hand, while a separate controller handled the rest of the body.
Tracking a task from live video
The biggest jump is in ER 2. The model watches a continuous video stream and checks in real time whether the task is going to plan — it detects errors, corrects motion, and recognizes when the job is done. In pinpointing a specific moment in footage it reaches 91.3% accuracy with a 0.96-second mean error, which DeepMind describes as the sub-second latency needed for safe operation. Rating task progress across five levels (0–100%) is harder — accuracy there drops to 57.4%.
Against the previous ER 1.6, the new model wins on tool orchestration in all three control modes and on safety and human-proximity benchmarks. DeepMind says it does so "at a fraction of the compute cost and 4x the execution speed".
Manipulation: where the model is strong, and where it isn't
The manipulation results show the real state of the art, unsweetened. There is a clear spread depending on the task and the gripper used.
| Task | Platform / hand | Success |
|---|---|---|
| Pick from shelf | Apollo 2 + Inspire | 76.3% |
| Pick from table | Apollo 2 + Inspire | 68.4% |
| Pick from floor | Apollo 2 + Inspire | 45.7% |
| Unscrew a bulb | Apollo 2 + Inspire | 92% |
| Insertion | Franka Duo | 89.6% |
On tasks that need full finger dexterity the model leads mainly in one narrow job — unscrewing a bulb — while the rest stay in the 32–44% range. The classic gripper fares best on the Franka Duo platform.
Multi-robot coordination is the launch's second pillar. DeepMind showed the Apollo 2 humanoid and a Franka F3 Duo working together on one process — a task a single machine could not do alone. Spatial reasoning now runs on raw video, and instrument reading covers 10 gauge types, including digital displays, rulers and thermometers.
Availability
Access is tiered. ER 2 is public in Google AI Studio and via the Gemini API, and in private preview on the Gemini Enterprise Agent Platform. The VLA models and On-Device 2 go, for now, only to early-access partners. On-Device 2 stands out for fast adaptation to a new robot — per DeepMind, a few hours and typically fewer than 200 examples are enough to port the model to a different build.
Why it matters
Moving from arm control to whole-body control widens the model's scope — one system now joins locomotion and manipulation that previously needed separate layers. Just as important is closing the perception loop: a model that watches its own actions on video and judges progress can correct errors without a human in the loop. Yet the middling manipulation scores outside narrow tasks show that a general-purpose humanoid remains a research goal, not a finished product.
What's next?
- The VLA and On-Device 2 models stay in early access — broader availability depends on DeepMind, with no date given.
- ER 2 is already testable in Google AI Studio and via the Gemini API, with code examples on GitHub.
- The low progress-rating accuracy (57.4%) is an identified weak point — an area to improve in future iterations.
