Overview
- Gemini Robotics 2, which DeepMind announced on July 30, is a family of three models that separate high‑level embodied reasoning (ER) from vision‑language‑action (VLA) execution and a lightweight on‑device variant.
- DeepMind showed the same model checkpoint running on Apptronik’s Apollo 2 humanoid and a Franka arm and reported partner work with Boston Dynamics to illustrate a hardware‑agnostic approach.
- The company says the on‑device model can adapt to a new robot body with fewer than 200 demonstration passes and a few hours of data collection, aiming to reduce the time and data needed for new embodiments.
- DeepMind has exposed ER2 and VLA through the Gemini API, Google AI Studio, and private enterprise previews, but independent third‑party validation, production timelines, deployment costs, and real‑world safety metrics remain unreported.
- The release reflects a shift from end‑to‑end reinforcement learning to a layered architecture that could speed adoption by hardware partners while concentrating commercial value in the intelligence layer and changing how robots are built and sold.