Related video
Open original video →We focus on written briefings; video stays on the original platform.
Google DeepMind's latest AI model, Gemini Robotics 2, marks a significant leap in embodied AI by enabling whole-body control of humanoid robots. While the previous version focused on upper-body manipulation, the new model extends control from feet to fingertips, allowing robots to walk, crouch, stretch, and manipulate objects with greater dexterity. This advancement, announced on July 30, 2026, is demonstrated on Apptronik's Apollo 2 robot, which can now bend over to pick up a watering can or retrieve specific items from a shelf.
For startups building humanoid robots, the implications are profound. Gemini Robotics 2 reduces the need for custom programming for each new task. Instead, a single AI model can handle a wide range of whole-body motions, from sealing a Ziploc bag to unscrewing a lightbulb. This generalist approach could dramatically lower development costs and time-to-market for new robot applications. However, Google DeepMind notes that movement speed still needs improvement, and real-world reliability remains unproven.
The model also includes an upgraded embodied reasoning system, Gemini Robotics ER 2, which enhances a robot's ability to understand its environment and execute multi-step tasks over extended periods. This version now understands when tasks begin and end, enabling more autonomous operation. Critically, it supports collaboration between multiple robots of different types, as shown in a video where Apollo 2 directs a dual-arm robot to clean a garage.
For operators, this technology could unlock new commercial use cases in logistics, manufacturing, and home assistance. The ability to control five-fingered hands opens up tasks that require fine manipulation, such as tying a trash bag or handling delicate objects. However, the current speed limitations mean that high-throughput applications like warehouse picking may still require specialized robots.
Founders should watch for the open-source release of Gemini Robotics 2 or its integration into robot platforms. Google DeepMind has not announced pricing or availability, but the trend is clear: generalist robot brains are becoming viable. Startups that can leverage these models to create versatile robots may gain a competitive edge, while those relying on custom programming for each task risk obsolescence.
The key takeaway for builders is to start experimenting with whole-body control models now. Even if Gemini Robotics 2 is not yet production-ready, the direction is set. Investing in modular hardware that can be controlled by such models will pay off as the software matures. Speed and reliability will improve, but the foundational shift from task-specific to generalist control is already here.
Source: The Verge.