Google DeepMind Debuts Gemini Robotics 2 Models
02 Aug 2026
Google DeepMind announced on a Thursday the launch of Gemini Robotics 2, a new suite of models positioned as the intelligence layer powering the next generation of adaptable robots. The release includes three variants — Gemini Robotics 2, Gemini Robotics ER 2, and Gemini Robotics On-Device 2 — each targeting different aspects of robot perception, planning, and control.
What's new
At the core is Gemini Robotics 2, a vision-language-action (VLA) model that converts visual and language input directly into motor control. According to the report, this model unlocks:
- Whole-body control — coordinated motion from feet to fingertips, enabling walking, crouching, stretching, and object manipulation.
- Advanced dexterity — control of five-fingered hands for complex manipulation tasks.
- Multi-robot collaboration — coordination across multiple robots of different types working on shared tasks.
A companion model, Gemini Robotics ER 2, handles embodied reasoning and vision-language understanding. It's designed to let robots communicate with humans, interpret the physical world, and plan multi-step tasks lasting several minutes. The update also improves the model's ability to recognize when tasks begin and end during extended operations. On the safety side, ER 2 can detect nearby humans, trigger safety tool calls, and bring robots to a safe stop.
The third variant, Gemini Robotics On-Device 2, is optimized to run locally on robotic hardware rather than in the cloud. Per the report, it can adapt to new robot embodiments — different shapes, sensors, and degrees of freedom — with just a few hours of data, a notable speed-up for integrating new hardware.
The announcement includes a demonstration caption showing a humanoid robot performing whole-body movement powered by the new model, and the report notes that humanoid platforms such as Apptronik's Apollo 2 and Google's own dual-arm robot are affected by this update.
What's still unclear
The report flags several gaps. There's no detail on which specific robots beyond Apollo 2 and Google's dual-arm system have been tested with Gemini Robotics 2. Release timeline, availability, and pricing for developers or enterprises are not specified. There are also no benchmark comparisons against prior robotics models or competitors' systems.
Google DeepMind itself acknowledges that robots still need to improve movement speed — though the report offers no specifics on what this limitation means in practice or when it might be addressed. The presence of active safety mechanisms, like stopping near humans, also suggests these systems still require built-in safeguards for unstructured environments — an implicit sign that safety challenges remain unresolved.
Why founders should care
For founders building in robotics or adjacent hardware/software layers, this release likely signals a few shifts worth tracking:
- Platform expansion: The combination of on-device and cloud-based models suggests startups may soon have more flexible options for deploying robotic intelligence, whether latency-sensitive or cloud-dependent.
- Lower prototyping barriers: Faster adaptation to new robot embodiments could reduce the time and data needed to get new hardware working with capable AI models — a meaningful advantage for teams iterating on custom robots.
- New coordination use cases: Multi-robot collaboration features may open opportunities in fleet-based automation, logistics, or shared-task robotics that weren't previously practical.
- Persistent constraints: Movement speed limitations and the continued need for safety tooling suggest founders should still plan around human-robot interaction safeguards and performance ceilings rather than assume fully solved real-world responsiveness.
Overall, the update points to broader applicability for manipulation-heavy and multi-robot tasks, though the lack of pricing, availability, and benchmark data makes it hard to gauge how soon — or how cheaply — these capabilities will reach startup teams building on top of them.