Google DeepMind revealed Gemini Robotics 2 on Thursday, a leap forward for embodied AI that can direct a humanoid robot’s full range of motion. Unlike its predecessor, which limited itself to the robot’s upper limbs, the new model orchestrates motions from the feet up through the fingertips, opening the door to tasks that require whole‑body coordination.
In video demos, the system guided Apptronik’s Apollo 2 robot to bend over and scoop up a watering can, then to locate and extract a specific item from a shelf. Those actions illustrate the model’s ability to walk, crouch, stretch and manipulate objects with a level of fluidity previously unseen in commercial robotics.
The upgrade also brings refined dexterity. Gemini Robotics 2 can now control five‑fingered hands, enabling robots to seal a Ziploc bag, tie a trash bag knot or unscrew a light‑bulb. Such fine‑motor tasks have long been a stumbling block for autonomous machines, and the new capability suggests a path toward more household‑ready robots.
Enhanced reasoning, safety and collaboration
Alongside the motion controller, DeepMind launched Gemini Robotics ER 2, an embodied‑reasoning model that helps robots interpret their surroundings, process multi‑step instructions and recognize when a task begins and ends. The updated ER 2 performs better over extended periods, allowing robots to complete longer, more complex sequences without losing context.
ER 2 also supports collaborative work between different robot types. A showcase video features Apollo 2 directing a dual‑arm robot to sort tools into a bin while cleaning a garage, demonstrating coordinated effort across platforms.
Safety received a notable boost. DeepMind describes ER 2 as its “safest robotics model to date,” citing improved human‑presence detection and automatic safety stops if a person approaches too closely. The system can trigger safety tool calls and bring the robot to a halt, reducing the risk of accidental contact.
Finally, the company upgraded its Gemini Robotics On‑Device Model, enabling the AI to run locally on a robot without an internet connection. This on‑device capability adapts more quickly to new robot embodiments, even those with drastically different shapes, sensors or degrees of freedom, expanding the model’s applicability across diverse hardware.
DeepMind acknowledges that movement speed still lags behind human performance, but frames the current release as a critical step toward real‑world tasks that demand whole‑body coordination. The company’s announcements signal a broader push to embed advanced AI directly into the hardware that will one day share our homes and workplaces.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.