Google announces 'Gemini Robotics On-Device', a high-performance robot AI model that operates locally without an Internet connection
Google announced the Gemini Robotics On-Device , a VLA model for robots, on June 24, 2025. Gemini Robotics On-Device is lightweight and can be processed using a local GPU, and has higher performance than existing local VLA models.
Gemini Robotics On-Device brings AI to local robotic devices - Google DeepMind
Excited to release Gemini Robotics On-Device and bunch of goodies today
— Dhruv Shah @ RSS (@shahdhruv_) June 24, 2025
🍬 on-device VLA that you can run on a GPU
🍬 open-source MuJoCo sim (& benchmark) for bimanual dexterity
🍬 Broadening access to these models to academics and developers https://t.co/mSjXTLuOeu
Google released the VLA model ' Gemini Robotics ' in March 2025, which is based on Gemini 2.0 and adds the 'function to output motion'. The newly announced Gemini Robotics On-Device is a model optimized for local operation of Gemini Robotics, and is designed to operate a two-armed robot with minimal computing resources.
Gemini Robotics On-Device can run without an internet connection, making it ideal for tasks that require low latency. In addition, even though it is a small model that can run locally, it can perform high-precision operations such as 'understanding natural language instructions and opening a zipper on a bag.'
Below is a graph showing the benchmark results for 'Gemini Robotics (blue)', 'the existing highest performance on-device model (blue-green)', and 'Gemini Robotics On-Device (purple)'. Gemini Robotics On-Device showed performance that far exceeded that of existing on-device models.
The graph below compares the achievement rates of 'easy tasks' and 'difficult tasks.' It can be seen that Gemini Robotics On-Device has an overwhelmingly higher task achievement rate than existing models.
Gemini Robotics On-Device also supports fine-tuning of the model, allowing it to adapt to new situations with just 50 to 100 demonstrations. The graph below compares the task completion rates of 'Gemini Robotics (blue)', 'the existing best-performing on-device model (blue-green)', and 'Gemini Robotics On-Device (purple)' with fine-tuning.
Gemini Robotics On-Device was originally designed for Google's robot arm ' ALOHA ', but it can now perform well with other companies' models such as ' Franka Research 3 ' and ' Apollo '. The image below shows a Gemini Robotics On-Device equipped with Gemini Robotics On-Device performing the task of 'hanging a red spirit level on the wall'.
Equipped with Gemini Robotics On-Device, Apollo was able to perform the task of 'selecting objects as instructed from multiple objects lined up on a table and packing them into a bag' using only local processing.
Google has made the Gemini Robotics On-Device development kit available to some developers, and is calling on interested developers to sign up for a waitlist .
We're also launching the Gemini Robotics software development kit (SDK) to help developers fine-tune the model for their own applications, including by testing it in the MuJoCo physics simulator. 🌐 pic.twitter.com/w8uOFSnCXK
— Google DeepMind (@GoogleDeepMind) June 24, 2025
Related Posts:
