Gemini Robotics ER 2: powering robotics with video understanding, task orchestration, and multi-robot collaboration

Google DeepMind describes ER 2 as a high-level brain that plans multi-step work, streams video and hands motor control to a VLA or other APIs. It reports 57.4% accuracy on progress classification and 91.3% on moment-finding with 0.96-second mean error, plus a Boston Dynamics Spot demo.

The post also adds multi-robot collaboration and says the model is available through the Gemini API, AI Studio and a private preview of Gemini Enterprise Agent Platform. Safety claims include gains on instruction-following and human-proximity tests.

This is Google DeepMind’s own launch post, so benchmark numbers, demos and safety claims are vendor-reported. The useful distinction is architectural: reasoning is separated from motor control and evaluated on timing, not only spatial questions.

Signal Desk's read: knowing when a task is done is the harder robotics problem, and video-based progress tracking is a plausible attack on it. The remaining risk is that orchestration quality depends on the lower-level controllers the model calls.

Multi-robot collaboration is the more speculative claim. Shared semantics can coordinate unlike machines, but real factories will still need fallbacks when a handoff fails.

ER 2 upgrades ER 1.6 and integrates with the Gemini Live API for bidirectional streaming. Developers can declare VLAs or navigation APIs as tools and stream video, audio or text.