Google DeepMind Unveils Gemini Robotics On-Device for Offline Robot Control

Google DeepMind just dropped an exciting update in the robotics space—Gemini Robotics On-Device, a new language model that lets robots perform tasks without needing an internet connection. This builds on their earlier Gemini Robotics model (released in March) but with a key upgrade: local processing.

Developers can now fine-tune robot movements using natural language prompts, making it easier to adapt robots for different tasks. Google claims its performance is nearly on par with its cloud-based counterpart and outperforms other on-device models (though they didn’t specify which ones).

Real-World Robot Skills: From Laundry to Assembly Lines

In demos, robots running this model successfully:

• Unzipped bags
• Folded clothes
• Adapted to new objects (like assembling parts on an industrial belt)

Originally trained for ALOHA robots, the model was later adapted to work on:

• Franka FR3 (a bi-arm industrial robot)
• Apptronik’s Apollo humanoid

Gemini Robotics SDK: Training Robots with Demonstrations

Google also announced a Gemini Robotics SDK, allowing developers to train robots using 50-100 task demonstrations in the MuJoCo physics simulator. This could speed up robot learning for real-world applications.

The Bigger Picture: AI’s Push into Robotics

Google isn’t alone in this race:

• Nvidia is building foundation models for humanoids
• Hugging Face is working on open models—and actual robots
• RLWRLD (a Korean startup) is developing foundational models for robotics

The future of AI-powered robots is heating up—and it’s happening offline, on-device, and in real time.

Want more tech insights?
Join us at TechCrunch Disrupt in Boston, MA (July 15) for deep dives into AI, robotics, and venture trends. Save $200+ on All Stage passes and connect with leaders from Precursor Ventures, NEA, Index Ventures, and Underscore VC.

👉 REGISTER NOW