How does AI control robots?
Google DeepMind's new AI model, Gemini Robotics 2, can control a robot's entire body, enabling whole-body motions and complex tasks. This update allows humanoid robots to perform a wider range of actions, from walking and manipulating objects to sealing bags and unscrewing lightbulbs. The new model also supports better dexterity and safety features.
Google DeepMind's latest unveiling - the Gemini Robotics AI model's newest iteration - is a significant leap forward, enabling the control of entire humanoid robots. Previously, the focus was on the upper body, but now Gemini Robotics 2 supports whole-body motions, from toes to fingertips. This means robots can perform a broader range of actions - walking, crouching, stretching, and manipulating objects, for instance. Take Apptronik's Apollo 2 robot, which can now bend down to pick up a watering can or retrieve specific items from a shelf.
The implications are huge for AI practitioners and users, as this update allows robots to tackle more complex, real-world tasks that demand whole-body coordination. Google DeepMind acknowledges there's still room for improvement in movement speed, but this is a major step towards achieving more intricate tasks. And with Gemini Robotics 2, robots also gain better dexterity, courtesy of more complex, five-fingered hands - think sealing a Ziploc or tying a trash bag.
The update also brings improvements to Gemini Robotics ER, a vision-language model that helps robots analyze their surroundings and perform multi-step tasks. Gemini Robotics ER 2 is more adept at completing tasks over an extended period and can now discern when tasks begin and end. It also enables multiple robots of different types to work together seamlessly - like Apollo 2 instructing Google's dual-arm robot to put tools away in a bin while cleaning the garage. Notably, the company says Gemini Robotics ER 2 is its safest model yet, with enhanced detection of nearby humans and automatic safety stops when someone gets too close.
Source: The Verge
Related posts
Mistral Large 4 Arrives on Hugging Face With Cybersecurity Focus
Mistral Large 4, a 1-trillion-parameter multimodal mixture-of-experts model with 49 billion active parameters and a 1-million-token context window, launched on Hugging Face October 29, 2026. It scores 82 percent on the Artificial Analysis Cyber Index for reproducing and patching vulnerabilities, 93 percent on Cybench, and 93.3 percent on Lakera's prompt-injection benchmark. Released under the Mistral Ecosystem Open-Weight License, it supports day-zero inference via transformers, vLLM,
NVIDIA Releases Nemotron 3.5 Transcribe for Promptable Multilingual
NVIDIA released Nemotron 3.5 Transcribe, an open-weight model on Hugging Face and ModelScope (October 28, 2026) that transcribes 40 languages, translates 31 to/from English, and attributes speakers in 13 languages-all via natural language prompts. The 30-billion-parameter model activates only 4 billion per token, topping open-weight benchmarks with 5.33% English WER and 18.26% diarization error rate across 21 languages. Granary v2 fine-tuning toolkit accompanies it.
Google DeepMind Unveils Gemini 4 Carbon Model
Google DeepMind unveiled Gemini 4 Carbon on October 28, 2026, claiming it as their most capable model yet with state-of-the-art performance in cybersecurity and long-horizon reasoning. The company says Carbon tops the updated v5 benchmark suite, though the specific test wasn't fully named. Technical details, parameter counts, and architectural changes remain undisclosed, and no timeline was given for API, consumer app, or NotebookLM availability. Independent evaluations and public acce