Gemini Robotics-ER 1.6 release notes / what's new

Gemini Robotics-ER 1.6 release notes / what's new

Google DeepMind has introduced Gemini Robotics-ER 1.6, a reasoning-first model designed to improve "embodied reasoning"—the ability of robots to bridge digital intelligence and physical action. This update significantly enhances spatial reasoning, multi-view understanding, and task success detection, enabling higher levels of autonomy for physical agents.

High-Level Reasoning and Integration

Gemini Robotics-ER 1.6 serves as a high-level reasoning engine for robots. It is capable of executing complex tasks by natively calling various tools, including:

  • Google Search: For retrieving external information.
  • Vision-Language-Action (VLA) models: For translating reasoning into physical movement.
  • Third-party user-defined functions: For custom robotic integrations.

The model demonstrates significant performance improvements over Gemini Robotics-ER 1.5 and Gemini 3.0 Flash, particularly in pointing, counting, and success detection.

Spatial Reasoning through Pointing

Pointing is used as a foundational capability for spatial reasoning, allowing the model to perform several critical robotic functions:

  • Precision Object Detection and Counting: Identifying and counting specific items within a scene.
  • Relational Logic: Defining "from-to" relationships (e.g., moving an object to a specific location) and making comparisons, such as identifying the smallest item in a set.
  • Motion Reasoning: Mapping trajectories and identifying optimal grasp points.
  • Constraint Compliance: Reasoning through prompts such as "point to every object small enough to fit inside the blue cup."

Gemini Robotics-ER 1.6 uses points as intermediate steps to perform mathematical operations for better metric estimations and to count items more accurately. In comparative tests, it outperformed Gemini Robotics-ER 1.5 in identifying the correct number of tools and avoiding hallucinations of items not present in the image.

Autonomous Success Detection

Success detection allows a robot to determine if a task is finished, enabling it to decide whether to retry a failed attempt or move to the next step in a plan. Gemini Robotics-ER 1.6 improves multi-view reasoning, which is essential for handling occlusions and poor lighting by synthesizing information from multiple camera streams (e.g., combining an overhead feed with a wrist-mounted feed) to form a coherent understanding of the environment.

Instrument Reading via Agentic Vision

Developed in collaboration with Boston Dynamics, Gemini Robotics-ER 1.6 introduces the capability to read complex industrial instruments, such as circular pressure gauges, vertical level indicators, and digital readouts. This is particularly useful for facility inspection robots like Spot.

To achieve high accuracy, the model utilizes "agentic vision," which combines visual reasoning with code execution through a multi-step process:

  1. Zooming: The model zooms into an image to capture small details on a gauge.
  2. Estimation: It uses pointing and code execution to estimate proportions and intervals.
  3. Interpretation: It applies world knowledge to interpret the final reading, including units and decimal places.

"Capabilities like instrument reading and more reliable task reasoning will enable Spot to see, understand, and react to real-world challenges completely autonomously."

— Marco da Silva, Vice President and General Manager of Spot at Boston Dynamics

Safety and Constraint Adherence

Gemini Robotics-ER 1.6 is described as the safest robotics model to date from Google DeepMind, with improved compliance with Gemini safety policies on adversarial spatial reasoning tasks.

Key safety improvements include:

  • Physical Safety Constraints: Better adherence to constraints such as "don't handle liquids" or "don't pick up objects heavier than 20kg."
  • Hazard Identification: In tests based on real-life injury reports (Asimov benchmark), the model improved over Gemini 3.0 Flash by 6% in text scenarios and 10% in video scenarios regarding the accurate perception of injury risks.

Availability

Gemini Robotics-ER 1.6 is available to developers via the Gemini API and Google AI Studio. Google has also provided a developer Colab notebook with examples for configuring the model for embodied reasoning tasks.

Sources