Gemini Robotics ER 2: 映像理解・タスク統合・マルチロボット協調でロボティクスを強化Gemini Robotics ER 2: powering robotics with video understanding, task orchestration, and multi-robot collaboration

Google DeepMind has introduced Gemini Robotics ER 2, a new iteration of its embodied reasoning model designed to help robots perceive their surroundings, plan multi-step actions, and coordinate with other machines to solve real-world tasks. The announcement matters because it targets three capabilities that have long constrained practical robotics: understanding what a camera actually observes over time, translating a broad instruction into a concrete sequence of actions, and enabling more than one robot to work toward a shared goal. Together, these are described as a step change for robotic applications.

The "ER" in the name refers to embodied reasoning, the branch of the Gemini Robotics family focused on the higher-level thinking a robot needs before it moves. In this framing, the model acts less like a direct motor controller and more like a planner or "brain" that interprets a scene, reasons about what should happen next, and issues commands that lower-level systems carry out. ER 2 builds on this lineage with what Google DeepMind characterizes as significant advances in video understanding, tool orchestration, and multi-robot collaboration.

Video understanding is central to the update. Rather than reasoning over a single still image, the model appears to work across temporal visual streams, allowing it to track objects, follow changes in a scene, and maintain spatial awareness as a task unfolds. This kind of continuous perception is important for tasks where timing and sequence matter, such as watching an object move, recognizing when a step is complete, or adjusting to an environment that does not stay static. Stronger video reasoning is generally considered a prerequisite for robots operating in unstructured, everyday settings rather than fixed industrial cells.

The second pillar, tool orchestration, corresponds to the task integration emphasized in the model's positioning. Here the system can decompose a high-level goal into smaller steps and call on external tools or subsystems to accomplish each one. This mirrors a broader trend in agentic AI, where a language or multimodal model coordinates other software and hardware components rather than trying to do everything itself. For robotics, orchestration is likely what allows a single instruction to become a chain of grasping, moving, checking, and re-planning actions.

The third advance, multi-robot collaboration, extends this reasoning across more than one machine. Coordinating multiple robots introduces challenges around task allocation, shared understanding of the environment, and avoiding conflicts, and improvements here suggest the model can help distribute work among several agents. Google DeepMind frames all three capabilities as enabling robots to reason through and solve complex real-world tasks, though the excerpt does not specify particular hardware platforms, pricing, or regional availability.

For context, Gemini Robotics ER 2 sits within a rapidly moving field. Earlier Gemini Robotics work paired a vision-language-action model for direct control with an embodied reasoning variant for planning, itself descended from Google's earlier robotics research such as RT-1, RT-2, and PaLM-E, and from collaborative datasets like Open X-Embodiment. The overarching idea is to bring the general knowledge and reasoning of large multimodal models into physical systems, so that robots can generalize to tasks and objects they were not explicitly trained on.

The release also arrives amid intense competition in embodied AI. NVIDIA has promoted its Project GR00T foundation models and Isaac robotics tooling, while startups including Figure AI, Physical Intelligence, and Skild AI, alongside humanoid programs such as Tesla's Optimus, are pursuing general-purpose robot intelligence from different angles. Many of these efforts share the same bet: that progress in perception and reasoning, not just mechanics, is the key bottleneck to useful robots.

As with most models at this stage, real-world performance will depend on how the system behaves outside controlled demonstrations, and independent evaluation will be important before broad conclusions are drawn. Reliability, safety, and latency all matter when reasoning must translate into physical motion. Still, by concentrating on video understanding, task integration, and multi-robot coordination, Gemini Robotics ER 2 appears aimed at the practical gaps that separate impressive demos from dependable deployment, and it signals Google DeepMind's continued investment in bringing Gemini's multimodal reasoning into the physical world.