On July 30, 2026, Google DeepMind pushed the Gemini Robotics series from tabletop robotic arms to full-body humanoid robots. Rather than a single model upgrade, this release splits the entire stack 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 previous Gemini Robotics 1.x series focused on upper-body tabletop manipulation in static scenes. DeepMind now explicitly layers robot intelligence into three tiers:
- Gemini Robotics 2 (VLA): A vision-language-action backbone that translates camera feeds plus natural-language instructions directly into motor control signals. The focus is full-body control "from fingers to toes"—previously navigation and manipulation were handled by separate models; now the VLA handles the entire continuous action chain from walking to crouching to grasping. In a demo, Apptronik's Apollo 2 received the instruction "put the watering can into the green box on the lowest shelf," then walked over, picked it up, crouched, and placed the can in the box without switching subtasks.
- Gemini Robotics ER 2 (embodied reasoning brain): A VLM-based high-level planner that decomposes tasks lasting minutes to tens of minutes, tracks progress, coordinates multi-robot collaboration, and can call Google Search or custom functions. ER 2 now connects through the Gemini Live API with bidirectional streaming endpoints at sub-second latency, reasoning about the next step while the underlying VLA executes.
- Gemini Robotics On-Device 2: A local-inference version for low-latency or weak-network scenarios. Its selling point is fast transfer: with fewer than 200 demonstration samples and a few hours, the model can be adapted to a brand-new robot embodiment.
- Grasp success rate: 68.4% tabletop, 76.3% on shelves, 89.6% on precise insertion tasks. Still short of human-level, but enough 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 0.96s mean absolute timing error, executing 4x faster than the previous generation. This is the key capability for "enough, stop" feedback—e.g., recognizing when a cup is full while pouring coffee.
- Spatial VQA and gauge reading: Extended from round dials to digital displays, linear scales, rulers, and liquid thermometers, covering 10 gauge types.
- Apptronik: Apollo 2 humanoid, demonstrated whole-body coordination.
- Boston Dynamics: Atlas (humanoid) and Spot (quadruped); Spot demoed ER 2 calling the Spot SDK via the Live API for natural-language tasks like "fetch popcorn."
- Agile Robots: industrial arm scenarios.
- Franka: F3 Duo dual-arm platform, demoed collaborating with Apollo 2.
- Sharpa: SharpaWave 22-DoF tactile hand for fine manipulation like knot-tying and peeling plastic wrap; also part of the NVIDIA Isaac GR00T reference design hardware.
- DeepMind blog: 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 figures are direct:
ER 2's two headline metrics matter differently: progress classification determines whether a robot can recover from mid-task failures (missing tool, hitting an obstacle) without restarting; moment-finding determines whether it can switch tasks at the precise instant (e.g., letting go of a lightbulb the moment it's screwed in). These are capabilities VLA models alone couldn't deliver—ER 2 fills the gap by using a VLM as the brain.
A wide hardware partner network
DeepMind simultaneously announced a long list of hardware partners covering mainstream humanoid robot makers:
Along with the simultaneously 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 running, use the notebooks in DeepMind's google-gemini/robotics-samples repo—starting by wiring up Spot.
My take: this release marks DeepMind shifting the robotics race from "selling models" to "building an ecosystem"—three models, five hardware partners, and two benchmarks put the second half of 2026's embodied-AI competition into a full-stack-vs-full-stack arena alongside NVIDIA Cosmos 1X Neo and Figure Helix. Notably, ER 2's 91.3% moment-finding accuracy with 0.96s error is the closest publicly reported metric to industrial-assembly tempo, and industrial deployments in Shanghai, Tokyo, and Seoul are likely to trial it first.
References: