Google DeepMind debuts Gemini Robotics‑ER 1.6, touting sharper spatial and physical reasoning for robots

What’s new

DeepMind today introduced Gemini Robotics‑ER 1.6, a “reasoning‑first” model aimed at giving robots a better sense of the physical world — not just instructions, but context and judgement. Short version: it can point more precisely, count more accurately, and decide when a task is actually done. It has been reported that the model also unlocks a new capability called instrument reading, letting robots interpret complex gauges and sight glasses — a feature the team says emerged through close work with Boston Dynamics. Who hasn’t wanted a robot that can tell you whether a pump is over‑pressured without making a frantic call?

Benchmarks and capabilities

It has been reported that Gemini Robotics‑ER 1.6 shows significant improvements over Gemini Robotics‑ER 1.5 and Gemini 3.0 Flash on spatial and physical reasoning tasks such as pointing, counting, and success detection. DeepMind highlights multi‑view understanding and better use of points as intermediate reasoning steps — the model can point to multiple items, avoid hallucinating nonexistent objects, and aggregate points to improve estimates. The blog post includes benchmark comparisons and notes that instrument‑reading tests were run with agentic vision enabled (ER 1.5 doesn’t support that feature), so caveats apply.

Why it matters

Success detection and reliable pointing aren’t sexy on their own, but they’re crucial for real autonomy: knowing when a task is finished prevents endless retries and keeps workflows moving. The instrument‑reading angle is the emotional kernel here — tiny, precise perception that can mean the difference between a human supervisor checking a gauge and a robot maintaining an assembly line unaided. This fits a broader trend: LLMs are getting married to embodied perception, and the partner is increasingly competent.

Access and context

Gemini Robotics‑ER 1.6 is available to developers via the Gemini API and Google AI Studio, and DeepMind published a Colab with starter examples for embodied reasoning tasks. Expect early adopters in industrial automation and logistics to test the limits fast. But as always, demos and benchmarks are one thing; real‑world deployment — with occlusions, bad lighting, and messy edge cases — will be the true proving ground.

Sources: deepmind.google