Gemini Robotics 2 brings whole body intelligence to robots
Points and comments are a snapshot, not live.
Google DeepMind's Gemini Robotics 2 enables whole-body control, dexterity, and multi-robot collaboration.
Google DeepMind announced Gemini Robotics 2, a suite of three models: a vision-language-action model (VLA) for motor control, an embodied reasoning (ER) model for planning and safety, and an on-device VLA for local execution. The system can control humanoid robots like Apptronik's Apollo 2 for whole-body tasks, perform dexterous manipulation with five-fingered hands and standard grippers, and enable multiple robots to collaborate. The ER model is available in Google AI Studio. Safety is improved via the ASIMOV-Agentic benchmark for refusal and uncertainty handling.
Multi-robot teamwork and fast adaptation to new robot bodies (within hours) are also highlighted. The models require early-access signup for VLA and on-device versions.
What commenters are saying
Many commenters remain skeptical about practical utility, comparing the field to GPT-1 level progress. Several practitioners note that demos often cherry-pick results and that real-world reliability is far off: hardware (grippers vs. hands), safety around humans, and task generalization remain unsolved. One commenter points to Sunday Robotics' clothes-folding as a rare working real-world deployment. Others debate whether VLA/VLM architectures will ultimately succeed or if alternative approaches (e.g., Yann LeCun's JEPA) will be needed. A technician who has worked on these systems says progress is real but household deployment is years away.