Skip to main content
Live
Main content

Google DeepMind's Gemini Robotics 2 takes control of a humanoid from feet to fingertips

The update extends the model from upper-body manipulation to full-body motion, letting Apptronik's Apollo 2 walk, crouch and pick items off shelves.

Jaeden Schafer
Editor in Chief · · 4 min read
Google logo

Google DeepMind released Gemini Robotics 2 on July 30, 2026, extending its humanoid control model from upper-body-only manipulation to what the lab calls control of the "entire" robot, from feet to fingertips. The update lets machines like Apptronik's Apollo 2 walk, crouch, stretch and manipulate objects under a single policy, closing the gap between navigation and manipulation that has forced most humanoid demos to cheat with pre-scripted locomotion.

In demos shared alongside the announcement, Apollo 2 bends over to pick up a watering can and pulls specific items off a shelf, motions that require coordinating balance, gaze, arm reach and grip in one loop. Gemini Robotics 2 also drives more complex five-fingered hands, enabling Apollo 2 to seal a Ziploc bag, tie a trash bag and unscrew a lightbulb — the kind of soft, force-sensitive tasks that have historically defeated general-purpose robot policies.

Google DeepMind concedes the robots "have more to advance in movement speed," a rare on-the-record hedge from the lab. But it frames the release as "an important step towards the skills needed to complete more complex, real-world tasks that require whole-body coordination." That is the load-bearing claim: whole-body coordination under one model, rather than a stitched pipeline of walking controller plus arm controller plus grasp planner.

Key facts

  • 01Gemini Robotics 2 extends control from upper body only to whole-body motion, covering walking, crouching, stretching and object manipulation.
  • 02The model can now drive five-fingered hands to seal a Ziploc, tie a trash bag, and unscrew a lightbulb.
  • 03Gemini Robotics ER 2 understands when tasks begin and end and can coordinate multiple robots of different types on a single job.
  • 04Google DeepMind calls Gemini Robotics ER 2 its safest robotics model to date, with new logic to detect humans and stop the robot.
  • 05The update landed on July 30, 2026, weeks after Gemini Robotics ER 2's video-native reasoning release.

Alongside the main model, Google DeepMind updated Gemini Robotics ER, its embodied-reasoning vision-language layer that translates instructions into robot plans. Gemini Robotics ER 2 handles longer task horizons and "now understands when tasks begin and end," a capability the previous version lacked and one that matters for any autonomous shift in a warehouse or home.

ER 2 also coordinates heterogeneous fleets. In one demo, Apollo 2 gives instructions to Google's dual-arm bench robot to put tools into a bin while cleaning a garage — a humanoid acting as foreman for a stationary manipulator. That multi-embodiment orchestration is what turns a single robot demo into a viable pitch to a facilities operator running mixed hardware.

Safety is the pitch Google DeepMind is leading with on ER 2, calling it the lab's "safest robotics model to date." The model can "better detect when humans are nearby, trigger safety tool calls and bring the robot to a safe stop if someone approaches too closely." For humanoids meant to operate around people rather than behind cages, that layer is the difference between a research demo and a customer pilot.

The lab also refreshed its Gemini Robotics On-Device Model, the version that runs locally without an internet connection. The updated on-device model adapts to new embodiments faster, including robots with "drastically different shapes, sensors and degrees of freedom." That portability matters because the humanoid market is fragmenting: Apptronik, Figure, 1X, Unitree, Agility and half a dozen others are shipping wildly different hardware, and a model that only runs on one body is a model that only sells to one customer.

The release builds on Gemini Robotics ER 2's video-native reasoning launch earlier this cycle, which gave robots a stronger visual grounding for multi-step instructions. The pattern is now clear: Google DeepMind is layering an embodied-reasoning brain, a whole-body control policy and an on-device runtime, then licensing the stack to humanoid OEMs like Apptronik rather than building its own robot.

Related · from this week
Google DeepMind's Gemini Robotics 2 controls humanoids across dexterous tasks
Jaeden Schafer · 5 min read →

Skeptics will note what the announcement did not include: no benchmark scores, no success-rate numbers on the demonstrated tasks, no comparison against Figure's Helix or Tesla's Optimus policies, and no timeline for commercial deployment of Apollo 2 running Gemini Robotics 2 in a paying customer's facility. The video reel is impressive but curated, and "more to advance in movement speed" is the kind of hedge that usually means the robots are still slow enough to be uneconomic for most manual labor.

The competitive read is that Google DeepMind is trying to become the Android of humanoid robotics — the model layer that any OEM can adopt — while Tesla and Figure keep their stacks vertical. If Gemini Robotics 2's whole-body policy generalizes across the Apollo 2, Google's own dual-arm rig, and future on-device targets with different degrees of freedom, the lab has a genuine platform play. If it doesn't generalize, it's another impressive lab demo that ships one customer at a time. The next signal to watch is a paying pilot with published task success rates, not another sizzle reel.

ShareXLinkedInEmail
AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

  • ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat
  • Generate images & video with Sora, Veo, Ideogram
  • Compare any two models side by side
  • From $8.99/mo · 80+ models, all included
Try AI Boxaibox.ai
Trusted by 3,000+ teams
Got a tip?

Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.

Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.

AI Box Daily briefingFree · Daily · No fluff

Stay ahead of everyone in AI.

The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.

Loved by 10,000+ AI professionals
Free forever. Unsubscribe with one click.

The briefing read inside teams at

Keep reading

More from Models

Google logo
Models

Google DeepMind's Gemini Robotics 2 controls humanoids across dexterous tasks

The updated model fuses a vision-language brain with two action models to run robots like Apptronik's Apollo 2 on real chores.

Jaeden Schafer5 min read
Google logo
Models

Google DeepMind ships Gemini Robotics ER 2 as a video-native brain for robots

The new embodied reasoning model tracks task progress from live video, orchestrates other robots, and runs at 4x the speed of larger rivals.

Jaeden Schafer5 min read
Google logo
Models

Google ships Nano Banana 2 Lite and Gemini Omni Flash to developers

DeepMind's fastest image model generates in 4 seconds at $0.034 per 1K images; Omni Flash matches Veo 3.1 Fast at $0.10 per second of video.

Jaeden Schafer5 min read