On July 30, 2026, Google DeepMind pushed the Gemini Robotics series from "tabletop robotic arms" to "full-body humanoid robots." This was not just a model upgrade — the entire stack was split into three cooperating models: the VLA backbone Gemini Robotics 2, the embodied reasoning brain Gemini Robotics ER 2, and a local replica, Gemini Robotics On-Device 2.
How the Three Models Divide the Work
The earlier Gemini Robotics 1.x series focused on upper-body tabletop manipulation and static scenes. DeepMind now explicitly splits robot intelligence into three layers:
- Gemini Robotics 2 (VLA): A vision-language-action backbone that translates camera frames plus natural-language instructions directly into motor control signals. The emphasis is whole-body control "from fingers to toes" — navigation and manipulation used to be two separate stages, but now the VLA itself handles the entire continuous action chain from walking to squatting to grasping. In a demo, Apptronik's Apollo 2, given the instruction "put the watering can into the green box on the bottom shelf," walked over, picked it up, squatted, and placed it in the box without any sub-task switching.
- Gemini Robotics ER 2 (embodied reasoning brain): A high-level planner on the VLM side, responsible for decomposing tasks lasting minutes to tens of minutes, tracking progress, coordinating multi-robot collaboration, and calling Google Search or custom functions. ER 2 now connects via the Gemini Live API's bidirectional streaming endpoint with sub-second latency, reasoning about the next step while the low-level VLA executes.
- Gemini Robotics On-Device 2: A local inference version targeting low-latency or weak-network scenarios. Its selling point is fast transfer — fewer than 200 demonstrations and a few hours are enough to adapt the model to a brand-new robot body.
- Grasp success rate: 68.4% tabletop, 76.3% on shelves, 89.6% on precise insertion tasks. Still short of human-level, but sufficient for industrial assembly pilot runs.
- Task progress classification (5 tiers, 0–100%): ER 2 achieves 57.4% accuracy, outperforming the previous ER 1.6 and other frontier models.
- Moment-finding: 91.3% accuracy with 0.96 seconds average absolute timing error, executing 4× faster than the previous generation. This is the key capability for robots to use "enough, stop" as feedback — e.g., recognizing when a cup is full while pouring coffee.
- Spatial VQA and gauge reading: Expanded from circular dials to digital screens, linear scales, rulers, and liquid thermometers, covering 10 gauge types.
- Apptronik: Apollo 2 humanoid, demonstrating whole-body coordination.
- Boston Dynamics: Atlas (humanoid) and Spot (quadruped); Spot demos show ER 2 calling the Spot SDK via the Live API for natural-language commands like "fetch the popcorn."
- Agile Robots: Industrial arm scenarios.
- Franka: F3 Duo dual-arm platform, demonstrating collaboration with Apollo 2.
- Sharpa: SharpaWave 22-DoF tactile hand for fine manipulation like tying knots and peeling plastic wrap; also part of the NVIDIA Isaac GR00T reference design.
- DeepMind blog post: https://deepmind.google/blog/gemini-robotics-er-2-powering-robotics-with-video-understanding-task-orchestration-and-multi-robot-collaboration
- Model card: https://deepmind.google/models/model-cards/gemini-robotics-er-2/
- Safety report: https://storage.googleapis.com/deepmind-media/gemini-robotics/Gemini-Robotics-2-Safety.pdf
- Demo code: https://github.com/google-gemini/robotics-samples
The Numbers
DeepMind's benchmark numbers are straightforward:
ER 2's two metrics matter differently: progress classification determines whether a robot can recover from mid-task failures (missing tool, obstacle) without starting over; moment-finding determines whether it can switch tasks at exactly the right time (e.g., letting go of a lightbulb the instant it's tightened). These are things VLA models alone have never achieved — ER 2 fills the gap by using a VLM as the brain.
A Much Wider Hardware Partner Network
DeepMind simultaneously announced a long list of hardware partners, covering essentially all major humanoid robot makers:
Alongside the newly released ASIMOV-Agentic benchmark and Safety Technical Report, ER 2 also beats ER 1.6 on human-proximity awareness and safe instruction following. In short: the models themselves aren't the headline — DeepMind is emphasizing the "stack" and the "ecosystem."
Availability and My Take
ER 2 is now available to developers via the Gemini API and Google AI Studio, with the Gemini Enterprise Agent Platform in private preview; the VLA backbone and On-Device 2 remain with early partners. To get started, use the notebooks in DeepMind's google-gemini/robotics-samples repository — wiring up Spot first.
My take: this release marks DeepMind shifting the robotics race from "selling models" to "building an ecosystem" — three models plus five hardware partners plus two benchmarks equals a "full stack vs. full stack" contest for late 2026, on the same tier as NVIDIA Cosmos 1X Neo and Figure Helix. Notably, ER 2's 91.3% moment-finding with 0.96-second error is the closest publicly reported metric to industrial assembly cadence; industrial deployments in Shanghai, Tokyo, and Seoul should be first in line for trials.
References: