DeepMind Unveils Gemini Robotics 2 for Advanced Robot Control
- tech360.tv
- 19 hours ago
- 3 min read
Google DeepMind has introduced Gemini Robotics 2, its latest vision-language-action, or VLA, model. This update builds on the first version, which demonstrated multimodal understanding for real-world action. The new iteration incorporates intelligent whole-body control, refined dexterity, and multi-robot collaboration, according to The Robot Report.
Gemini Robotics 2 allows robots to process every movement, enabling a wider array of tasks. Humanoid robots could walk, crouch, stretch, and interact with objects to tidy a disordered room. The model also facilitates cooperation, permitting multiple robots to work together on tasks. The system runs directly on devices, adapting to new robot designs within hours, requiring under 200 data examples.
And, DeepMind announced two supplementary models. Gemini Robotics ER 2, an embodied reasoning (ER) model, functions as an agent, allowing robots to communicate with humans, understand physical environments, and plan multi-step tasks. ER 2 is accessible via Google AI Studio and a private preview on Gemini Enterprise Agent Platform.
The second model, Gemini Robotics On-Device 2, is an optimised VLA designed to operate locally on robotic hardware. This model achieves rapid adaptation to different robot forms. Early-access partners can utilise the VLA and On-Device models. DeepMind's previous models focused on controlling a humanoid's upper body for tabletop activities.
Gemini Robotics 2 now extends physical artificial intelligence to full-body movements. It provides control for entire humanoid robots, converting instructions into coordinated whole-body actions. The Apptronik Apollo 2 robot processed an instruction to place a watering can into a bin, executing the necessary sequence. But, DeepMind acknowledges robot movement speed requires further progress. This development advances capabilities for complex, real-world tasks demanding full body synchronisation.
Precision is key for useful robots. Gemini Robotics 2 introduces advanced physical dexterity across various end effectors. The model controls the five-fingered, 22-degree-of-freedom SharpaWave hand on the Apollo 2 robot for delicate actions. It also operates standard two-fingered parallel grippers on a Franka Duo platform, enabling intricate dexterous tasks such as compact packing.
Most real-world tasks involve multiple steps over an extended duration. So, Gemini Robotics ER 2 acts as the robot's primary processing unit, interpreting user instructions, facilitating human communication, and monitoring progress. It allows robots to perform multi-step tasks, correct errors, adapt to new scenarios, and identify key events. The organisation also implements multi-robot collaboration for complex workflows.
For situations with limited network access, Gemini Robotics On-Device 2 addresses requirements for operations without network latency or internet connection. DeepMind engineered this model for such restrictions. The On-Device model supports multiple embodiments inherently, using "motion transfer" techniques. It can adapt to new bi-arm robot designs in a few hours of training.
Google DeepMind maintains safety forms a core principle for robotics. Gemini Robotics 2 enhances robotic safety for unpredictable environments and human collaboration. DeepMind introduced ASIMOV-Agentic, a new benchmark to evaluate agentic safety orchestration and uncertainty resolution. And, with its enhanced embodied reasoning, Gemini Robotics ER 2 is presented as DeepMind's safest robotics model to date, improving human proximity detection and safety shutdowns.
Gemini Robotics 2 offers intelligent whole-body control, advanced dexterity, and multi-robot collaboration.
It operates locally on devices, adapting to new robot designs rapidly with minimal data.
DeepMind also released Gemini Robotics ER 2 for embodied reasoning and Gemini Robotics On-Device 2 for local operation.
Safety measures include the ASIMOV-Agentic benchmark and enhanced human proximity detection in ER 2.
Source: The Robot Report
