Robotics · 3 min read

Feet to Fingertips: Gemini Robotics 2 Brings Whole-Body Control to Humanoids

Google DeepMind's Gemini Robotics 2 controls humanoids from feet to fingertips, adds multi-robot collaboration, and ships the ASIMOV-Agentic safety benchmark.

By Classy AI News · July 31, 2026

Feet to Fingertips: Gemini Robotics 2 Brings Whole-Body Control to Humanoids

Tabletop manipulation was only half the problem. A robot that can grasp a mug but cannot walk to the shelf, crouch to the bottom row, and place the mug precisely still fails most household chores.

On July 30, 2026, Google DeepMind introduced Gemini Robotics 2 — a vision-language-action stack that extends from prior upper-body control to whole-body humanoid coordination, from feet to fingertips, alongside refreshed embodied-reasoning and on-device models.

Industrial robotic arm operating in a factory environment

Three models, one physical stack

The release bundles three components:

  • Gemini Robotics 2 (VLA): converts vision and language into motor commands; now controls full humanoids and bi-arm platforms with advanced dexterity on hands and grippers
  • Gemini Robotics ER 2: high-level embodied reasoning for multi-minute tasks, human communication, and multi-robot coordination — available on Google AI Studio and in private preview on Gemini Enterprise
  • Gemini Robotics On-Device 2: low-latency local VLA that adapts to new robot embodiments in a few hours with under 200 examples

Carolina Parada, who authored DeepMind's announcement, framed the advance as teaching robots to "reason through every movement" — walking, crouching, stretching, and manipulating in cluttered spaces.

Apollo 2 walks, picks, and places

Demonstration videos show Apptronik's Apollo 2 humanoid responding to natural-language instructions like "put the watering can into the green bin in the bottom shelf." The robot walks to a table, picks up the can, navigates to shelving, and places it — whole-body coordination previous Gemini Robotics releases did not attempt.

DeepMind notes movement speed still needs improvement. The milestone is capability breadth, not production-ready cycle time.

Dexterity gains extend to 22-degree-of-freedom SharpaWave hands on Apollo 2 — tying knots, sealing Ziploc bags — and parallel grippers on Franka Duo platforms for tight packing tasks.

Engineer working with automation equipment in a workshop

Multi-robot collaboration and on-device adaptation

Gemini Robotics ER 2 now tracks when tasks begin and end, executes longer sequences involving hundreds of decisions, and coordinates heterogeneous robots via shared semantic understanding.

One demo shows Apollo 2 directing a dual-arm Franka platform to collect tools while cleaning a garage — handoffs across embodiments that single-robot stacks struggle to orchestrate.

On-device adaptation inherits motion-transfer techniques from Gemini Robotics 1.5, enabling few-hour retargeting to new shapes, sensors, and degrees of freedom — relevant for warehouses and homes where cloud latency is unacceptable.

Safety: ASIMOV-Agentic benchmark

Physical capability without safety orchestration is a liability. DeepMind introduced ASIMOV-Agentic, a benchmark evaluating whether embodied reasoning agents refuse unsafe tool calls from VLAs, assess task feasibility, and request human help when uncertain.

Gemini Robotics ER 2 is described as DeepMind's safest robotics model to date on safety-instruction-following and human-proximity tests — halting when people approach and resuming when clear.

The full ASIMOV-Agentic benchmark is published on Hugging Face; a safety technical report accompanies the release.

Close-up of robotic components and wiring on a workbench

Private preview, public reasoning

VLA and on-device models remain in early-access partner programs; ER 2 is the most broadly accessible piece today. DeepMind positions Gemini Robotics 2 as progress toward "physical AGI" — generalist intelligence that adapts across robot bodies rather than single-task automation.

Partners listed include Apptronik, Boston Dynamics, and Agile Robots.

For the robotics industry, July 30 marks the moment a major frontier lab treated locomotion and manipulation as a single model problem — not a pipeline stitched from separate controllers.

Sources

Newsletter

Get the dispatch

One field. One email when we publish. Privacy.