Google DeepMind Unveils Gemini Robotics 2

Since the release of version 1.5, almost a year has passed. This is quite a long break by today's standards. Let's see what Google has achieved during this time.

The most important aspect is the VLA model, which now controls the entire humanoid body from toes to fingertips, rather than just the upper body. The hands have fine motor skills with 22 degrees of freedom. It cannot be said that the robot is super agile (the success rate on multi-finger tasks is around 30-40%), but this is progress.

On top of the VLA model, there is also the Gemini Robotics ER 2 orchestrator. This is a separate generation model, and here it is simply a VLM. That is, if the VLA outputs ready motor commands, then ER operates at a higher level, performing reasoning, breaking tasks into subtasks, recognizing objects and their coordinates, and other "intelligent" work.

This model has also been significantly enhanced and has a killer feature of coordinating the work of several different types of robots to accomplish a single task. Additionally, ER can now intervene in the actions of VLA at any stage of task execution, which is critical for self-correction.

Another model introduced in this generation is Gemini Robotics On-Device 2 – a lightweight VLA for local inference.

The main distinction of Gemini Robotics is its versatility. The models are trained on heterogeneous multi-embodiment data, allowing them to work on various types of robots without tuning. The GR2 On-Device takes this property to a practical level: you can take this model and, in just a few hours and about 200 demonstrations, make it work well on any new body (with custom sensors, DoF, kinematics, etc.).

Additionally, in a fun twist, Google has created the first benchmark for assessing robot safety and named it ASIMOV-Agentic. They naturally claimed their model as "the safest in the world" according to this test.