THG News · AI & Research

Google DeepMind Releases Gemini Robotics ER 2 for Robot Planning and Coordination

· Updated

Google DeepMind released Gemini Robotics ER 2 on July 30, 2026, an embodied reasoning model that Google describes as a high-level brain for robots. According to a post on Google's blog by Steven Hansen and Peng Xu, the model is publicly available to developers through the Gemini API and Google AI Studio, and in private preview on the Gemini Enterprise Agent Platform. Google says it is a significant upgrade over Gemini Robotics ER 1.6.

Embodied reasoning models of this kind sit above the controllers that move a robot. According to Google, ER 2 integrates with vision-language-action (VLA) models and robotics APIs, planning multi-step tasks and handing the motion itself to those systems.

Video understanding and progress tracking

Google says ER 2 can watch a robot's video feed to judge how far a task has progressed and whether it succeeded or failed, so the robot can correct itself. On a progress classification test that sorts task completion into five levels, ER 2 scored 57.4% accuracy, according to Google. On a test of locating specific moments in video, Google reported 91.3% accuracy with a mean absolute distance of 0.96 seconds. The company also said the model can read more than 10 types of instruments and offers improved visual question answering.

Orchestration and multiple robots

Google said ER 2 can orchestrate multi-step tasks through the Gemini Live API, and showed a demonstration built with Boston Dynamics' Spot. The company also demonstrated ER 2 coordinating different robots on a shared workflow, including Apptronik's Apollo 2 humanoid and the Franka F3 Duo.

On safety, Google said ER 2 can halt a humanoid robot when a person comes nearby and resume work once the area is clear. It said the model outperforms ER 1.6 on its Safety Instruction Following and Human Proximity benchmarks. Google also published code examples on GitHub for developers.

The wider second-generation family

Google DeepMind's Gemini Robotics model page now lists three second-generation models: Gemini Robotics 2, a VLA that the page says is available to try through Google AI Studio; Gemini Robotics ER 2; and Gemini Robotics On-Device 2, a lighter VLA built to run locally on robot hardware.

The On-Device 2 model card, dated July 2026, says access is limited to a select group of trusted testers and that the model builds on Gemini Robotics 1.5 and Gemma on-device models. It is designed for bi-arm robots, and the card notes limitations on high degree-of-freedom systems and mobile platforms. On robot platforms introduced after training, the card reports 53.3% success on SO101 against 6.7% for the first version, and 75.6% against 33.3% on Dexmate platforms, after extended training.

For research labs and integrators, the split matters: ER 2 is broadly available through the API, while the on-robot VLA remains in a limited testing program, according to Google DeepMind.

Sources