Gemini Robotics-ER 1.6: Enhanced Embodied Reasoning

Gemini Robotics-ER 1.6 is an embodied reasoning model designed to help robots understand their physical environment more accurately and act more autonomously. It handles high-level judgments such as spatial reasoning, task planning, and success detection, and can directly call Google Search, VLA models, and external custom functions.

Compared to the previous version, 1.5, and Gemini 3.0 Flash, performance in spatial and physical reasoning—such as pointing, counting, and success detection—has significantly improved. Developers can access it via the Gemini API and Google AI Studio, and Colab examples are also provided.

The core feature is Pointing.

• Object detection and counting
• Relational logic, such as finding the smallest item in a set
• Trajectory estimation and identifying optimal grasp points
• Handling constraints such as "point to an object small enough to fit in the blue cup"

Pointing is used as an intermediate step to solve complex tasks step by step. In experiments, the model accurately identified hammers, scissors, paintbrushes, and pliers, and hallucinations—such as pointing to a nonexistent wheelbarrow or drill—were reduced. 3.0 Flash showed comparable performance but had a weakness in recognizing pliers.

Success Detection has also been strengthened. Through multi-view reasoning that interprets multiple camera viewpoints together, the model more reliably determines when a task is complete even amid occlusion, lighting changes, or ambiguous instructions. For example, it can verify across multiple video feeds whether a task like "put the blue pen into the black pen holder" has been completed.

The newly added Instrument Reading feature is useful in real industrial settings. When Boston Dynamics Spot captures images of thermometers, pressure gauges, sight glasses, and digital instruments, the model comprehensively interprets liquid levels, scale markings, unit text, and even multiple needles. It sequentially leverages zooming, pointing, and code execution to attempt sub-tick-level precision readings.

Safety has also improved. Compliance with Gemini's safety policies has increased, and constraints such as "do not handle liquids" or "do not lift objects over 20kg" are now reflected at the pointing stage. In text and video safety scenario evaluations based on real injury reports, it also scored higher than Gemini 3.0 Flash.

Google DeepMind is collaborating with the robotics community to collect failure cases, and by receiving 10 to 50 labeled images from each application domain, aims to improve robustness in the next release. Gemini Robotics-ER 1.6 is currently available to try in Google AI Studio.