THG News · AI & Research
Google DeepMind Releases Gemini Robotics 1.5 — Open Weights for First Time
Google DeepMind released Gemini Robotics 1.5 on June 12, marking the first time the company has made a flagship robotics foundation model available with open weights for non-commercial research use.
What's New
- On-device inference at 30 Hz on a single NVIDIA Jetson Thor
- Cross-embodiment transfer: a single policy works across Aloha, Franka, Apptronik Apollo, and Unitree G1
- Long-horizon reasoning: the model can decompose multi-step tasks (e.g., "pack a lunchbox") without per-step prompts
- Safety constraints baked into the planning layer via constitutional-AI style critics
Benchmark Results
On the new BEHAVIOR-1K benchmark, Gemini Robotics 1.5 scores 71% task success on previously unseen household tasks — a 23-point improvement over the 1.0 model and roughly twice the score of the open-source OpenVLA-2.
Open Weights Surprise
The decision to release weights — albeit under a research-only license — is widely seen as a competitive response to Hugging Face's open LeRobot SmolVLA-3 release in May. DeepMind says a fully commercial-licensed Gemini Robotics 1.5 Pro will follow in Q4 via Google Cloud.
Adoption
Apptronik, Boston Dynamics, and Sanctuary AI all confirmed they are evaluating Gemini Robotics 1.5 as a fallback policy. NVIDIA announced day-one optimizations in its Isaac platform.