Google Unveils Gemini Robotics On-Device Model for Local Robot Control
New Gemini Model Brings Advanced Robotics Capabilities Without Internet Dependency
Google DeepMind has launched a new AI language model called Gemini Robotics On-Device, enabling robots to perform complex tasks locally—without requiring an internet connection.
- Building on the cloud-based Gemini Robotics model released in March, this new on-device version lets robots operate in real time and respond to natural language prompts for customized control.
Performance and Versatility
Google claims that Gemini Robotics On-Device performs nearly on par with its cloud-based predecessor and outperforms other on-device models in general benchmarks—though specific competing models were not named.
- In demonstrations, the model powered robots to unzip bags, fold clothes, and adapt to new scenarios, such as assembly on industrial belts using the bi-arm Franka FR3 robot and Apollo humanoid robot by Apptronik.
- Originally trained on ALOHA robots, the model showcased strong adaptability across various robotic platforms.
Empowering Developers with Gemini Robotics SDK
Google also announced the Gemini Robotics SDK, enabling developers to easily train robots on new tasks by showing them 50 to 100 demonstrations—all on the MuJoCo physics simulator.
- The SDK allows fine-tuning and control using natural language, simplifying integration for research and commercial use.
The Growing AI Robotics Ecosystem
The release comes amid rising activity in the AI-robotics space:
- Nvidia is developing a platform for foundation models for humanoids.
- Hugging Face is working on open-source models, datasets, and robotics hardware.
- Korean startup RLWRLD, backed by Mirae Asset, is building foundational models for robotics.
