Gemini Robotics ER 2 Ships: Multi-Robot Orchestration, 91.3% Moment-Finding, and a Safety Layer Above the VLA

A high-level brain for physical agents

Google DeepMind launched Gemini Robotics ER 2 on July 30, 2026 as part of the broader Gemini Robotics 2 suite — an embodied reasoning model that acts as a robot's planning layer while lower-level vision-language-action models handle motor execution. The model is now publicly available via the Gemini API, Google AI Studio, and private preview on the Gemini Enterprise Agent Platform.

Think of ER 2 as the agent that observes continuous video, reasons about multi-step tasks lasting several minutes, coordinates tool calls including Google Search, and — in this release — enables multi-robot collaboration across heterogeneous hardware.

Benchmarks Google published

Google DeepMind reports measurable gains over Gemini Robotics ER 1.6:

• 57.4% accuracy on progress-classification tasks
• 91.3% accuracy on moment-finding — identifying the precise video frame when a key event occurs
• 0.96 seconds mean absolute distance in moment-finding evaluations
• Four times the execution speed of larger model categories on these tasks

The model card states Gemini Robotics ER 2 is based on Gemini 3.5 Flash and accepts text, image, video, and audio with a 131,072-token input limit and 65,536-token output limit.

Multi-robot collaboration in practice

No single robot fits every task. Google demonstrated Gemini Robotics 2 enabling Apptronik's Apollo 2 and a Franka F3 Duo to collaborate via shared semantic understanding — handing off subtasks a single machine could not complete alone.

ER 2 tracks its own progress, adapts when steps fail, and knows when to advance. That progress understanding is new: the model can pinpoint when tasks begin and end, not just what to do next.

Safety: ASIMOV-Agentic and human proximity

Each Gemini Robotics 2 release includes expanded safety evaluation. Google introduced ASIMOV-Agentic, a benchmark for agentic safety orchestration measuring whether the reasoning agent refuses unsafe tool calls from a VLA and requests human intervention when uncertain.

Gemini Robotics ER 2 also advances human-proximity monitoring — detecting nearby people, triggering safety tool calls, and stopping humanoid motion when someone enters the workspace. Google states it outperforms ER 1.6 and other evaluated frontier models on safety instruction-following benchmarks.

The model card explicitly warns against safety-critical applications and lists healthcare and transportation among prohibited high-stakes settings.

Where this sits in the stack

Gemini Robotics 2 is a three-model family: the VLA handles whole-body humanoid control and dexterous manipulation; ER 2 plans and orchestrates; On-Device 2 runs locally with fast adaptation to new embodiments in a few hours from under 200 examples.

ER 2 is the piece enterprises can call today through standard API channels. The VLA and on-device models remain early-access partner releases.

Sources

• Google DeepMind — Gemini Robotics 2 brings whole body intelligence to robots (July 30, 2026)
• Google — Introducing Gemini Robotics ER 2 (July 30, 2026)
• Google DeepMind — Gemini Robotics ER 2 Model Card (July 2026)
• Google AI for Developers — Gemini Robotics ER 2 Preview API docs (July 2026)