- Published on
Powering robotics with video understanding, task orchestration, and multi-robot collaboration
- Authors
- Name
- DeepHorus
Google has unveiled Gemini Robotics ER 2, a new embodied reasoning model designed to give robots a high-level ‘brain’ for real-time interaction and multi-step planning. The model, announced as the successor to Gemini Robotics ER 1.6, lets robots chat with humans, understand physical spaces, and hand off motor control to lower-level vision-language-action (VLA) systems. It also natively supports external tools such as Google Search and custom functions, enabling robots to think and act without disruptive pauses. This launch marks a significant step toward practical physical AI agents capable of working alongside people in everyday environments.
The new model addresses a critical limitation in robotics: spatial reasoning alone is insufficient for dynamic tasks. Gemini Robotics ER 2 emphasizes temporal intelligence, allowing robots to track their own progress from continuous video feeds and adapt when something goes wrong. It introduces progress classification, assigning each video frame into five completion levels (0-20%, 20-40%, 40-60%, 60-80%, 80-100%), and moment finding to verify that tasks like tightening a light bulb or tying a trash bag are fully completed. These capabilities help robots know exactly when to move on, a notoriously hard challenge in automation.
In benchmark evaluations, Gemini Robotics ER 2 consistently outperformed ER 1.6 in tool orchestration across three control modes: real VLA, simulated VLA, and human teleoperation. The model is integrated into the Gemini Live API through a bidirectional streaming endpoint optimized for low-latency tasks. This integration eliminates the jarring ‘stop-and-think’ behavior common in earlier systems, enabling fluid orchestration of action models and robotics APIs. Google demonstrated the approach using a Boston Dynamics Spot robot, which fetched a popcorn snack on natural language command, with the code available on GitHub.
Developers can access Gemini Robotics ER 2 today through the Gemini API, Google AI Studio, and a private preview on the Gemini Enterprise Agent Platform. The public availability is expected to accelerate experimentation in physical AI, allowing startups and research labs to build more adaptive robots. Industry observers see this as a competitive move against other foundation models targeting robotics, though Google did not disclose specific commercial partnerships beyond Boston Dynamics. The model’s ability to generalize to novel situations and self-correct in shared spaces suggests near-term uses in warehouse logistics, home assistance, and remote teleoperation.
Looking ahead, Google plans to iterate on multi-robot collaboration, which the current release introduces in a preliminary form. Enabling robots to work together in shared spaces could unlock workflows beyond the capacity of any single unit. As developers test the tool-orchestration workflows and share feedback, future versions may refine the balance between high-level reasoning and real-time motor control. For now, Gemini Robotics ER 2 offers a practical glimpse into how embodied reasoning can bridge the gap between language instructions and reliable physical action.
