This week, Google DeepMind revealed Gemini Robotics 2, an intelligence layer comprising three new models to power more adaptable physical AI. Together, Google says these models will give robots more dexterous, full-body control to work together and complete a wide range of multi-step tasks. The vision-language-action (VLA) model, Gemini Robotics 2, debuted on Thursday, and converts vision and language inputs into motor control so robots can flex from feet to fingertips with enough dexterity in hands and grippers to complete delicate tasks, like closing a Ziploc bag. The lightweight version, Gemini Robotics On-Device 2, runs locally so robotic applications can keep running even without internet connectivity. Meanwhile, the embodied reasoning (ER) model, Gemini Robotics ER 2, lets robots understand their surroundings and communicate with humans so they can devise plans to carry out multi-step tasks — think emptying a dishwasher and putting items away. Chris Matthieu, VP of the developer ecosystem at RealSense, tells The New Stack this is what it’ll take for robots to graduate from simple, isolated actions to real-world physical assistance: “The hard part isn’t making the first decision — it’s recovering from the hundredth when the world has changed. Doors are closed, objects get moved, people walk into the scene, batteries drain, and sensors become partially occluded.”Full-body control and greater dexterityAccording to Google, combining VLA and ER models means humanoids can go further to complete a range of tasks, literally. Where its previous Gemini Robotics 1.5 model could only control a robot’s upper body to perform tabletop tasks, Gemini Robotics 2 brings intelligence to the entire humanoid body, enabling it to walk, crouch, stretch, and handle various objects. That means developers could create systems where robots are capable of executing commands like, “fetch the book from the top shelf.” “Humans make this look effortless because our visual cortex constantly feeds our motor cortex with spatial information. Physical AI needs that same continuous perception loop.”With full-body control also comes finer motor skills. Referencing the five-fingered, 22-degree-of-freedom SharpaWave hand on Apptronik’s Apollo 2 humanoid robot, Google says its Gemini Robotics 2 enables the robot to perform intricate tasks, like tying a knot, with the finesse and dexterity of a human hand.Source: GoogleWhat may seem like simple tasks for humans, Matthieu tells The New Stack that these capabilities represent a profound leap for humanoids as every subsystem must become interdependent to avoid obstacles, localize itself, and coordinate both arms and legs:“Humans make this look effortless because our visual cortex constantly feeds our motor cortex with spatial information. Physical AI needs that same continuous perception loop.”Multi-step, minutes-long task execution – from one robot or a whole groupWhen asked what’s been holding back the next breakthrough in robotics, Matthieu points to the unpredictability of the real world, telling The New Stack: “Long-horizon autonomy requires continuous replanning based on real-time perception rather than blindly following an original plan.”Now, Google is trying to help robots find their way with its ER model, which it calls “the robot’s high-level brain.” It can communicate with people, process task instructions, and then coordinate with the VLA model to carry out actions until completion — even those that last several minutes and require hundreds of decisions. Notably, the ER model also enables robots to self-correct if something goes awry. Plus, Gemini Robotics 2 should allow developers to quickly adapt to different robot embodiments, requiring as few as 200 examples and just a couple of hours of adaptation time for new bi-arm robot embodiments. That means skills can be adapted to new systems, even those with different shapes, sensors, and degrees of freedom. And if one intelligent, highly dexterous humanoid isn’t enough to do the job on its own, Google says Gemini Robotics 2 also allows it to collaborate with other robots to carry out more complex workflows.Local intelligence and a new safety benchmark While Gemini Robotics 2 and Gemini Robotics ER 2 seem to do most of the heavy lifting, the lightweight version of the VLA model shouldn’t be overlooked. Actually, Matthieu tells The New Stack that’s where he sees a keen advantage for privacy, reliability, and determinism, noting, “A robot shouldn’t have to wait on an Internet connection before deciding whether to stop, avoid a person, or grasp an object.”Eventually, he predicts robotic systems will adopt a hybrid architecture, where perception, motion planning, and safety decisions take place at the edge and cloud models handle deeper reasoning and long-term learning.“A robot shouldn’t have to wait on an Internet connection before deciding whether to stop, avoid a person, or grasp an object.”Of course, keeping a fleet of intelligent humanoids in check will require more controls. Google also introduces ASIMOV-Agentic, a new benchmark for agentic safety orchestration and uncertainty resolution that can measure the agent’s ability to refuse unsafe tool calls and flag a human for help when it’s not sure it can complete a task safely. Gemini Robotics ER 2 is now available on Google AI Studio and in private preview on Gemini Enterprise Agent Platform. Both VLA models are available to early-access partners. The post Gemini Robotics 2 brings us one step closer to physical AGI appeared first on The New Stack.