Google DeepMind has developed Gemini Robotics, a pair of AI models designed to bring sophisticated reasoning and action capabilities to robots. Built on the Gemini foundation models, these systems combine vision, language, and motor control to enable multi-step, general-purpose physical tasks.
Gemini Robotics consists of two complementary models:
- Gemini Robotics-ER 1.5 (Embodied Reasoning, ER) – a vision-language model (VLM) optimized for planning and reasoning in physical environments. It interprets visual and textual input, creates multi-step task plans, and can natively call digital tools like Google Search or third-party APIs to gather relevant data. The ER model acts as the high-level planner, generating natural language instructions that guide the robot through complex sequences.
- Gemini Robotics 1.5 (Vision-Language-Action, VLA) – a vision-language-action model that converts ER-generated instructions into precise motor commands. Unlike traditional VLA models, it incorporates an internal reasoning loop, allowing the robot to “think” about each step, segment complex tasks, and adjust actions based on environmental feedback.
The combined system enables multi-level task reasoning. For example, when sorting objects into bins based on local recycling guidelines, the ER model generates a step-by-step plan including data retrieval, object classification, and action sequencing. Gemini Robotics 1.5 then executes the plan, analyzing each movement, adjusting grip and trajectory, and reporting progress in natural language for transparency.
A key innovation is cross-embodiment learning. Motion strategies learned on one robot – such as the two-armed Aloha 2 – can transfer to other platforms, including humanoid robots like Apollo or the bi-arm Franka, without specialized retraining. This capability accelerates development, allowing new robots to inherit prior knowledge and generalize skills to new tasks.
Gemini Robotics-ER 1.5 achieves state-of-the-art performance on 15 academic embodied reasoning benchmarks, including Embodied Reasoning Question Answering (ERQA), Point-Bench, RefSpatial, RoboSpatial-VQA, and Where2Place. Its high performance spans pointing, image-based question answering, video understanding, and trajectory prediction, demonstrating advanced spatial reasoning and task progress estimation.
DeepMind has integrated semantic and physical safety mechanisms into both models. High-level reasoning considers task safety before execution, while onboard collision avoidance ensures operational safety. The upgraded ASIMOV benchmark provides improved tail coverage, annotations, and video modalities for evaluating semantic safety, confirming the models’ ability to respect both environmental and human-centric constraints.
By combining reasoning, planning, tool use, and action generalization, Gemini Robotics enable robots to perform complex, multi-step tasks autonomously. Gemini Robotics-ER 1.5 is available via Google AI Studio for developers, while Gemini Robotics 1.5 is currently accessible to select partners, paving the way for advanced research and practical deployment of intelligent robotic agents.
