Gemini Robotics On-Device brings AI to local robotic devices

Source: DeepMind Date: 2025-06-24 URL: https://deepmind.google/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/

Summary

Google DeepMind released Gemini Robotics On-Device, an on-robot VLA model that runs without network connectivity, enabling low-latency dexterous manipulation. It’s the first Gemini Robotics model available for fine-tuning — adapting to new tasks with 50–100 demonstrations. Cross-embodiment transfer demonstrated: trained on ALOHA but deployed on Franka FR3 and Apollo humanoid. Available through trusted tester program with an SDK and MuJoCo simulator integration.

Implications

Fine-tunable on-device VLA is the developer unlock. The first-to-fine-tune positioning is more important than on-device latency. Robotics startups and research labs need to customize manipulation policies to their specific hardware and task profiles. A fine-tunable Gemini model with 50–100 demo sample efficiency is a practical tool, not just a research capability.

Cross-embodiment transfer at 3 platforms. ALOHA → Franka → Apollo is a real cross-embodiment test, not a same-robot transfer. That’s harder and more useful — it means the model has learned manipulation representations rather than robot-specific motor programs. The implication is that fine-tuning generalizes better than training from scratch per embodiment.

Network-independence is the industrial requirement. Factory floors, warehouses, and field robotics often have unreliable connectivity. On-device inference addresses the core enterprise deployment blocker for cloud-dependent robotics AI. This changes Gemini Robotics from “lab demo” to “deployable in production” status.

Watch:

• Fine-tuning results in independent robotics labs — does 50–100 demo sample efficiency hold on novel tasks?
• When trusted tester program opens to broader developers
• Whether CFS, Boston Dynamics, or other named Gemini Robotics partners adopt the On-Device model