Skip to content
NewsModel release

Google DeepMind ships Gemini Robotics 2 with whole-body humanoid control and on-device adaptation

· by Pondero Newsdesk

The short version

Google DeepMind released three Gemini Robotics 2 models on July 30, 2026, extending AI control from upper-body-only to full humanoid motion, adding multi-robot collaboration, and delivering an on-device variant that adapts to new robot bodies in hours with fewer than 200 examples.

Google DeepMind ships Gemini Robotics 2 with whole-body humanoid control and on-device adaptation

Google DeepMind's previous Gemini Robotics models handled a robot's upper body for tabletop tasks. On July 30, the lab shipped three simultaneous models that extend AI control to legs, torso, and fine manipulation at once, and add an on-device variant that adapts to an entirely new robot body in a few hours.

What happened

Google DeepMind released three models on July 30, 2026 under the Gemini Robotics 2 name. Per the DeepMind blog, the flagship Gemini Robotics 2 is a vision-language-action model that converts visual and language input directly into motor control. Where its predecessors controlled only a humanoid's upper body, this model manages the full robot from feet to fingertips, including walking to a location, crouching to reach a lower shelf, and placing objects at a specific destination.

Benchmarks from the release show the model running on Apptronik's Apollo 2 humanoid with Inspire hands: 68.4% success picking items off a table, 45.7% picking from the floor, and 76.3% picking from a shelf. On a two-arm Franka Duo platform using a gripper, the model reached 74.2% on general pick-and-place, 78.9% on diverse tool kitting, and 89.6% on precise insertion tasks. Multi-finger dexterity on the SharpaWave hand showed a wider range: bulb unscrewing reached 92%, while ziplock sealing came in at 40% and dustpan tasks at 32%, per the DeepMind benchmarks.

The second model, Gemini Robotics ER 2, is an embodied reasoning model that acts as the planning layer above the VLA. It accepts natural-language instructions, decomposes multi-step tasks spanning several minutes and hundreds of decisions, tracks whether each step succeeds, and coordinates across multiple robot types simultaneously. Multi-robot collaboration is new in this generation. ER 2 is available now on Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, per the announcement.

The third model, Gemini Robotics On-Device 2, runs entirely on the robot's hardware without a network connection. It adapts to new robot bodies, including platforms with different sensors, actuators, and degrees of freedom, with fewer than 200 examples and a few hours of adaptation time, per DeepMind. During testing it ran on Dexmate, SO101, and Trossen platforms in addition to the primary Apollo 2 and Franka Duo hardware.

Why it matters

The prior generation of Gemini Robotics reached into a scene from a fixed position. Full-body control changes the viable task space to environments built around human movement: stepping between rooms, reaching high or low shelves, working in cluttered spaces. The benchmark spread tells the real story more plainly than a single number would. Precise insertion at 89.6% and floor pickup at 45.7% reflect a model that handles constrained, structured subtasks well but still struggles with the kind of unstructured motion people take for granted.

The on-device adaptation capability carries sharper near-term stakes for operators. Industrial and field deployments often cannot route all sensor data through a cloud inference endpoint, whether because of latency, connectivity constraints, or security requirements. A model that adapts to a new robot body in hours using fewer than 200 examples removes a significant bottleneck from deployment planning. Previously, adapting a trained robot policy to a new hardware platform took days or weeks of additional data collection and fine-tuning.

For developers, ER 2 being live on Google AI Studio now means robotics reasoning pipelines can be prototyped immediately without hardware. The model accepts camera input and language instructions, reasons about task state, and coordinates robot actions through tool calls to a VLA, making it the entry point for teams that want to experiment before signing up for physical robot access. The Gemini Enterprise Agent Platform preview provides a path to production-grade deployment for larger organizations.

Google DeepMind also published a new safety benchmark alongside this release. ASIMOV-Agentic measures whether an embodied reasoning agent can refuse unsafe tool calls from a VLA, predict when a task is not achievable, and proactively request human intervention under uncertainty. ER 2 improved over prior versions on safety-constraint following and human-proximity detection benchmarks, per the accompanying Gemini Robotics 2 Safety Technical Report. That proximity-detection improvement matters for markets where collaborative-robot safety standards require a machine to stop when a person enters its operating zone.

Context

The Gemini Robotics line has moved through several generations in 18 months: the original in March 2025, version 1.5 in September 2025, ER 1.6 in April 2026, and now the Gemini Robotics 2 family in July 2026. Acknowledged hardware partners for this release include Apptronik, Boston Dynamics, and Agile Robots. The VLA and on-device models remain behind an early-access signup; ER 2 is the only member of the family accessible without a formal application as of the July 30 announcement. Physical Intelligence has released its pi-0 and pi-0.5 VLA models as the nearest direct competitor, though that lab has not published a simultaneous three-tier release of VLA, reasoning, and on-device models from a single announcement.

What to watch next

Apptronik, Boston Dynamics, and Agile Robots have not announced production deployment schedules for the Gemini Robotics 2 VLA or On-Device 2 as of July 30. Watch whether the Gemini Enterprise Agent Platform preview opens broadly in the next 30 to 60 days. Physical Intelligence and Tesla Optimus have not yet responded publicly, and any benchmark releases from those teams will clarify how the competitive landscape sits in the second half of 2026.

Sources