Humanoid hardware is advancing quickly, but a useful robot also needs to understand what is happening, decide what to do and notice when its plan has failed. Google DeepMind’s Gemini Robotics 2 family is designed to connect those steps.
Google announced Gemini Robotics 2 in 2026 as a vision-language-action model capable of controlling whole humanoid bodies, including walking, reaching, balancing and fine manipulation. A related model, Gemini Robotics ER 2, focuses on video understanding, task orchestration and coordination between multiple robots. (Google DeepMind: Gemini Robotics 2 whole-body intelligence; Google DeepMind: Gemini Robotics ER 2.)
The demonstrations are a meaningful research advance, but they are not a retail product review. Most examples come from Google and its trusted testing partners in controlled settings, and Google’s model card lists limitations and safety considerations. Household readiness cannot be inferred from a successful video. (Google DeepMind: Gemini Robotics 2 whole-body intelligence; Google DeepMind: Gemini Robotics ER 2 model card.)
How Gemini Robotics 2 turns instructions into movement
A language model produces words. A robot model must connect words and images to a body with particular joints, cameras and reach. If asked to place a watering can on a lower shelf, it needs to identify the object, plan a route, bend without falling, grasp without crushing and release at the right position. (Google DeepMind: Gemini Robotics 2 whole-body intelligence.)
Google says Gemini Robotics 2 can control an entire Apptronik Apollo 2 humanoid through such multi-stage tasks. Earlier systems often focused on tabletop arms or upper bodies. Whole-body control expands the problem from hand movement to simultaneous locomotion, balance and manipulation. (Google DeepMind: Gemini Robotics 2 whole-body intelligence.)
What Gemini Robotics ER 2 adds to robot reasoning
Gemini Robotics ER 2 is a high-level embodied-reasoning model based on Gemini 3.5 Flash. It accepts combinations of text, images, video and audio, then helps interpret progress, choose tools or robot skills and decide when a plan needs correction. Lower-level control models still execute the physical movement. (Google DeepMind: Gemini Robotics ER 2; Google DeepMind: Gemini Robotics ER 2 model card.)
This division resembles a supervisor and specialist. The reasoning model can break a goal into tasks and monitor the scene, while a dedicated controller handles the rapid motor commands. That architecture can also assign different parts of a job to different robots. (Google DeepMind: Gemini Robotics ER 2.)
Why video understanding matters
Robots need to recognise not only objects but events: whether a drawer opened, a cup slipped or a person interrupted the task. Google reports 91.3% accuracy on its moment-finding evaluation, with a mean timing error of 0.96 seconds, and says the model runs fast enough for sub-second physical decision loops. (Google DeepMind: Gemini Robotics ER 2.)
These are company-reported benchmark results, not independent evidence of performance across all homes or factories. Benchmark construction, camera placement and task selection affect the result. Still, accurately locating the moment when something changed is a useful capability for self-correction and safety monitoring. (Google DeepMind: Gemini Robotics ER 2; Google DeepMind: Gemini Robotics ER 2 model card.)
One model, several robot bodies
Robotics teams usually train around a specific arm or body. A motion that works on one machine may not fit another robot’s dimensions, gripper or joint limits. Google is trying to make skills more transferable by training across multiple platforms and pairing shared reasoning with body-specific control. (Google DeepMind: Gemini Robotics 2 whole-body intelligence; NVIDIA Isaac GR00T robotics foundation-model overview.)
Transfer matters economically. If every robot needs a separate intelligence stack, software development fragments across hardware brands. A more general model could let manufacturers focus on bodies and tasks while reusing some learned concepts—although safe deployment will still require body-specific testing. (Google DeepMind: Gemini Robotics ER 2 model card; NVIDIA Isaac GR00T robotics foundation-model overview.)
The safety problem is physical
An incorrect chatbot answer can mislead. An incorrect robot action can drop an object, damage equipment or injure someone. Google says its robotics work uses layered safety measures, including semantic checks, low-level collision controls and model evaluations. The model card also warns that performance may degrade outside tested conditions. (Google DeepMind: Gemini Robotics ER 2 model card.)
A robust system needs more than a policy that says ‘be safe’. It requires mechanical limits, emergency stops, constrained operating areas, human oversight and testing of rare failures. The closer a robot works to people, the more the complete system matters—not only the intelligence model. (Google DeepMind: Gemini Robotics ER 2 model card; International Federation of Robotics: humanoids—vision and reality.)
What would count as convincing progress?
The next evidence should measure success over long deployments: unfamiliar objects, clutter, changing light, interruptions and repeated shifts. Useful reporting would include how often humans intervene, how quickly the robot recovers and whether safety performance holds as tasks become less scripted. (Google DeepMind: Gemini Robotics ER 2 model card; International Federation of Robotics: humanoids—vision and reality.)
Gemini Robotics 2 points toward a future in which a robot can understand a goal rather than merely replay a sequence. The technology is moving from isolated movements toward perception, planning and teamwork. Whether it becomes a dependable ‘robot brain’ will be decided in messy environments, not polished demonstrations. (Google DeepMind: Gemini Robotics 2 whole-body intelligence; Google DeepMind: Gemini Robotics ER 2.)
Reporting note: This article separates demonstrated results from company forecasts and staged demonstrations. It is general information, not purchasing, employment or investment advice.
Explore more evidence-led coverage in Robotics & Humanoids.
The hardware poses its own challenges. Read why robot hands are difficult to build for a closer look at the touch sensing and grip control that manipulation requires.
Leave a Reply