researchMulti-modelProprietary
Gemini Robotics
by Google DeepMind
Architecture
Multi-model
Parameters
Undisclosed
Training Data
50+ robot fleet, Internet-scale pretraining
License
Proprietary
Status
RESEARCH
Open Source
No
Robots Supported
Everyday RobotsAtlas (Boston Dynamics)Various research arms
About
Google DeepMind's multi-model robotics system combining RT-2 for vision-language-action reasoning, AutoRT for autonomous data collection, and SARA-RT for efficient transformer inference. Leverages the full Gemini model family for high-level task understanding.
Key Differentiator
Multi-model pipeline leveraging Gemini LLMs with specialized robotics modules (RT-2, AutoRT, SARA-RT)
Funding Context
Google DeepMind internal; Boston Dynamics partnership via Alphabet ownership
Milestones
2023-07RT-2 model published with web-scale pretraining
2024-01AutoRT autonomous data collection fleet deployed
2024-03SARA-RT efficient inference announced
2025-03Gemini Robotics ER announced with Boston Dynamics integration
2025-10Atlas electric running Gemini-powered manipulation tasks