Google Gemini Robotics 2: Whole-Body Humanoid Robot Brain Released

Google DeepMind released the Google Gemini Robotics 2 suite on July 30, 2026 — introducing a new humanoid robot brain and expanding the capabilities of google deepmind robotics with three new models.

This update represents a major step forward for gemini robotics. The suite consists of three distinct models. Gemini Robotics 2 (VLA) is a vision language action model (VLA model) that outputs motor commands directly to a robot's joints under a single policy. Gemini Robotics ER 2 (embodied reasoning) acts as the high-level planning layer: it watches video, interprets instructions, plans multi-step task sequences, and can call Google Search mid-task. Gemini Robotics On-Device 2 is a lightweight version optimized to run locally on hardware.

Access is split: ER 2 is publicly available in Google AI Studio and via the Gemini API (free rate-limited tier plus pay-as-you-go). The VLA and On-Device 2 models are restricted to trusted hardware partners. The primary demo platform is Apptronik's Apollo 2 humanoid. Weights are not released — all three models are API-only.

Google also released ASIMOV-Agentic alongside the models, a new safety benchmark measuring an agent's ability to refuse unsafe tool calls, recognise infeasible tasks, and escalate to humans.