Gemini Robotics 1.5 brings AI agents into the physical world

Source: DeepMind Date: 2025-10-23 URL: https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/

Summary

Google DeepMind launched two complementary robotics models: Gemini Robotics 1.5, a VLA that “thinks before acting” and transfers skills across different robot bodies; and Gemini Robotics-ER 1.5, a VLM for high-level planning that integrates Google Search and achieves SOTA on 15 academic datasets including ERQA and Point-Bench. ASIMOV v2 was released alongside. ER 1.5 is available on the Gemini API; Robotics 1.5 is restricted to select research partners.

Implications

The VLA/VLM split is an architectural bet on cognitive division of labor. Robotics 1.5 (VLA) handles low-level perception-to-action; ER 1.5 (VLM) handles high-level planning and tool use. The two-tier architecture mirrors how humans separate reactive motor control from deliberate reasoning — and makes each component independently improvable. Cross-embodiment transfer in the VLA means skills learned on one robot shape transfer without full retraining on a second body, which is the bottleneck that killed most previous sim-to-real pipelines.

Google Search integration in ER 1.5 is the distinguishing detail. A robot that can query the web to resolve unfamiliar objects or procedures has a qualitatively different capability ceiling than one limited to training-time knowledge. This is the agentic robotics pattern — the robot is an agent that uses tools, not a fixed policy.

SOTA on 15 datasets including ERQA and Point-Bench signals capability breadth, not cherry-picking. ERQA tests embodied reasoning across diverse manipulation tasks; Point-Bench tests spatial grounding. Sweeping both, plus 13 other benchmarks, suggests general visuomotor capability rather than narrow task specialization.

Watch:

• When Gemini Robotics 1.5 exits restricted partner access — the gap between ER 1.5 (API-available) and Robotics 1.5 (restricted) is where the real deployment moat lives
• Cross-embodiment transfer claims: which robot bodies and which skill classes transfer cleanly?
• ASIMOV v2 dataset adoption by non-Google robotics labs as a training benchmark