Google DeepMind's Gemini Robotics 2 gives humanoid robots full-body control — from feet to fingertips
The upgraded model unifies locomotion and manipulation in a single inference, runs on hardware from Apptronik, Boston Dynamics, and Agile Robots, and pairs with a publicly available reasoning model — but Google's own benchmarks show success rates as low as 32%.
What matters
- Gemini Robotics 2 unifies locomotion and manipulation in a single model inference, eliminating the handoff between a pre-programmed walking stack and a separate AI manipulation model — a first for the Gemini Robotics line.
- The release includes three sub-models: the core VLA model for whole-body control (private preview, waitlist), Gemini Robotics ER 2 for embodied reasoning (publicly available via Gemini Live API and Google AI Studio), and Gemini Robotics On-Device 2 for local inference without a network connection.
- Google's own published benchmarks show per-task success rates ranging from 32% (sweeping with a dustpan) to 92% (light-bulb unscrewing), with floor-level object picking at 45.7%, general whole-body manipulation at 45.7%–76.3%, and gripper kitting at 74.2%–89.6% — well below what the demo videos suggest.
- ER 2 is built on Gemini 3.5 Flash with a 128K-token context window, reasons over continuous video, classifies task progress into five completion bands, reads 10 instrument types, can resume from the last successful step after errors, and can call Google Search, navigation APIs, or custom functions as tools mid-task.
- The VLA model was demonstrated on Apptronik Apollo 2 (with SharpaWave and Inspire hands plus a Franka Duo/Robotiq gripper setup), Boston Dynamics Atlas, and Agile Robots hardware. DeepMind also released ASIMOV-Agentic, a safety benchmark on Hugging Face under CC-BY-4.0. Alphabet shares rose approximately 3% in pre-market trading on July 31.
Open source
- Repository:
- https://huggingface.co/ (ASIMOV-Agentic safety benchmark)
- License:
- CC-BY-4.0
Launch facts
- Availability:
- ER 2 publicly available via Gemini Live API and Google AI Studio; VLA and On-Device models in private preview via waitlist
- Platforms:
- Apptronik Apollo 2, Boston Dynamics Atlas, Agile Robots, Franka Duo dual arm with Robotiq gripper
Benchmarks
| Benchmark | Gemini Robotics 2 (VLA) | Source |
|---|---|---|
| Light-bulb unscrewing | 92% | vendor-reported |
| Gripper insertion and kitting tasks | 74.2%–89.6% | vendor-reported |
| General whole-body manipulation | 45.7%–76.3% | vendor-reported |
| Picking an object off the floor | 45.7% | vendor-reported |
| Sweeping with a dustpan | 32% | vendor-reported |
| Multi-finger dexterity (range across tasks) | 32%–92% | vendor-reported |
Numbers come from the linked sources; vendor-reported results are labeled and worth independent verification.
What happened
On July 30, 2026, Google DeepMind released Gemini Robotics 2, a family of three AI models designed to give humanoid robots unified whole-body control — coordinating walking, crouching, balancing, and fine finger manipulation through a single model inference rather than separate locomotion and manipulation stacks.
The flagship Gemini Robotics 2 is a vision-language-action (VLA) model that converts visual and language inputs directly into motor commands across an entire humanoid body, from feet to fingertips. Previous Gemini Robotics versions, including 1.5 released in September 2025, controlled only the upper body while relying on pre-programmed locomotion systems for walking. Collapsing that split into a single end-to-end policy eliminates the engineering headaches of timing handoffs between two codebases and debugging failures at the seam.
The release also includes Gemini Robotics ER 2, an embodied-reasoning vision-language model that serves as a high-level planner. ER 2 reasons over continuous video, plans multi-step tasks lasting several minutes, coordinates multiple robots on shared missions, and can call Google Search, navigation APIs, or developer-declared custom functions as tools mid-task. It also classifies task progress into five completion bands and can resume from the last successful step after an error. ER 2 is built on Gemini 3.5 Flash with a 128K-token context window and is publicly available now via the Gemini Live API and Google AI Studio.
A third model, Gemini Robotics On-Device 2, is a smaller variant designed to run locally on robotic hardware without a network connection.
The VLA model was demonstrated on Apptronik's Apollo 2 humanoid (configured with SharpaWave and Inspire hands plus a Franka Duo/Robotiq gripper setup), Boston Dynamics' Atlas, and Agile Robots systems. DeepMind also released ASIMOV-Agentic, a safety benchmark available on Hugging Face under CC-BY-4.0.
Why it matters
The architectural shift matters because a robot that walks to a shelf while reasoning about where to place its hand, maintaining balance as the load shifts, is closer to what a factory or warehouse actually needs. Separate walking and grasping systems are each impressive; combining them into one learned policy is the prerequisite for general-purpose humanoid labor.
But the most honest part of the launch is the benchmark chart Google published alongside the demo videos. Per-task success rates range widely: 92% on light-bulb unscrewing and up to 89.6% on gripper kitting tasks, but only 45.7% for picking an object off the floor, 45.7%–76.3% for general whole-body manipulation, and as low as 32% for sweeping with a dustpan. Multi-finger dexterity remains the weakest axis. DeepMind itself acknowledges that the robots "have more to advance in movement speed."
In other words, the demo reel shows a humanoid smoothly carrying a watering can across a room; the fine print says that same floor-picking task succeeds less than half the time. The gap between the two is where the real engineering work remains.
Alphabet shares rose approximately 3% in pre-market trading on July 31 following the announcement, suggesting investors see commercial potential even at current reliability levels.
What to watch
- Whether the unified VLA model's success rates improve meaningfully as more training data accumulates, particularly on the hardest tasks (floor-level manipulation, multi-finger dexterity).
- How broadly ER 2's public API exposes its multi-robot collaboration and spatial logic capabilities versus what stays gated in the private-preview stack.
- Which additional humanoid hardware companies partner with DeepMind beyond Apptronik, Boston Dynamics, and Agile Robots — access to a state-of-the-art unified control model could become a key differentiator.
- Whether the ASIMOV-Agentic safety benchmark gains traction as a community standard for evaluating agentic robot safety.
- Pricing, SLA, and commercial deployment timelines, none of which have been disclosed.
What to do next
Developers
Try Gemini Robotics ER 2 via the Gemini Live API or Google AI Studio today, and join the waitlist for early access to the core Gemini Robotics 2 VLA model on DeepMind's model page. Also explore the ASIMOV-Agentic safety benchmark on Hugging Face.
ER 2 is publicly available now for reasoning, multi-step task planning, tool calling, and error-resume behavior, while the VLA model for whole-body control requires waitlist access. Experimenting with ER 2 provides a head start on DeepMind's robotics stack, and ASIMOV-Agentic offers a community safety evaluation tool.
Founders
Assess whether a unified whole-body control model reduces the software burden for your robotics application, particularly if you are building on Apptronik, Boston Dynamics, or Agile Robots hardware.
If Gemini Robotics 2 proves reliable, it could replace custom locomotion-plus-manipulation stacks with a single pre-built control and reasoning layer, lowering engineering costs and accelerating time to market.
PMs
Map use cases that require both locomotion and dexterity — warehouse picking, delivery, eldercare, or facility maintenance — and identify where unified whole-body control unlocks new product features, while benchmarking against the published 32%–92% success-rate range to set realistic deployment thresholds.
The shift from upper-body to whole-body control expands the task space humanoid robots can address, but Google's own benchmarks show current success rates vary widely and are well below what most production environments require for the hardest tasks.
Investors
Track which humanoid hardware companies partner with DeepMind for Gemini Robotics 2 integration, as access to a state-of-the-art unified control model could be a key differentiator. Note Alphabet's ~3% pre-market move on July 31 as an early market signal.
Control software is a critical layer in humanoid robotics; Apptronik, Boston Dynamics, and Agile Robots are already demonstrated, and the positive stock reaction suggests the market sees commercial potential. Additional partnerships may signal market leaders.
Operators
Begin scoping pilot environments where humanoid robots with full-body control could supplement human labor, focusing on tasks requiring both movement and fine dexterity in dynamic spaces — but plan for current success rates well below operational thresholds.
Whole-body control with dynamic balancing is a prerequisite for robots navigating real-world environments like warehouses or factories while performing manipulation tasks, but the published benchmarks indicate significant reliability gaps remain.
How to test
- 1Visit the Gemini Robotics ER 2 model page at deepmind.google/models/gemini-robotics/gemini-robotics-er/ and click 'Try Gemini Robotics ER 2' to access it via the Gemini Live API, or open it in Google AI Studio.
- 2Test ER 2's embodied reasoning capabilities: multi-step task planning, spatial logic, progress classification across five completion bands, and multi-robot collaboration scenarios.
- 3Test ER 2's error-recovery behavior: interrupt a multi-step task mid-execution and verify whether it resumes from the last successful step.
- 4Test ER 2's tool-calling features: have it call Google Search mid-task, invoke a navigation API, or use a developer-declared custom function.
- 5Test ER 2's instrument-reading capabilities with digital displays, linear scales, rulers, or liquid thermometers across the 10 supported instrument types.
- 6Visit the Gemini Robotics 2 VLA model page at deepmind.google/models/gemini-robotics/vla/ and click 'Join waitlist for early access'.
- 7If granted early access to the VLA model, integrate it with a supported humanoid hardware platform per DeepMind's documentation.
- 8Test whole-body tasks that combine locomotion and manipulation (e.g., walk to a target, crouch, pick up an object, carry it to a new location).
- 9Evaluate multi-robot collaboration scenarios using ER 2 for task planning and delegation, with the VLA model handling motor execution.
- 10Download the ASIMOV-Agentic safety benchmark from Hugging Face and run it against your robot's agentic decision-making to evaluate safety compliance.
Caveats
- Gemini Robotics 2 VLA is in private preview; waitlist enrollment does not guarantee access or a timeline.
- Gemini Robotics ER 2 is publicly available via the Gemini Live API and Google AI Studio, but it's unclear how much of the multi-robot collaboration and spatial logic is exposed through the public API versus the private-preview stack.
- Google's own published benchmarks show success rates ranging from 32% to 92% depending on the task — real-world reliability in unpredictable settings is currently far below operational thresholds for the hardest tasks.
- DeepMind acknowledges robots 'have more to advance in movement speed.'
- Supported hardware platforms (Apptronik Apollo 2, Boston Dynamics Atlas, Agile Robots) are not generally available for purchase or testing by individual developers.
- No pricing, SLA, or commercial deployment timeline has been disclosed.