Gemini Robotics (2025)

Σχετικά με Google DeepMind

Ιδρύθηκε 2010 · London, UK / Mountain View, USA

Google DeepMind is the merged research organisation of DeepMind (London, 2010) and Google Brain, responsible for the Gemini frontier-model family. The DeepMind robotics group has a long lineage of generalist-robot work, from SayCan and RT-1 through RT-2 and the Open-X-Embodiment collaboration. In March 2025 the team announced Gemini Robotics, an advanced Vision-Language-Action (VLA) model built on top of Gemini 2.0 that brings multimodal reasoning, web knowledge and code into low-level robot control. Gemini Robotics is positioned as a generalist foundation model for dexterous bimanual manipulation, and is being trialled with hardware partners including Apptronik (Apollo humanoid), Agility (Digit) and Boston Dynamics. The model is research-only and not publicly available, but accompanies a public tech report and demo gallery.

Αρχιτεκτονική

Vision-Language-Action (VLA) transformer on top of Gemini 2.0 multimodal foundation

Gemini Robotics is a Vision-Language-Action model built from Gemini 2.0 by adding an action decoder that converts multimodal context (camera images, language instructions, optional robot state) into continuous low-level control commands. The backbone retains Gemini's web-scale multimodal pretraining (text, image, video, code), so the policy inherits broad world knowledge and language understanding. On top of that, the model is fine-tuned on a large multi-embodiment robot demonstration corpus including Aloha 2 bimanual setups, third-party humanoids (Apptronik Apollo, Agility Digit) and Google's own manipulation platforms. It outputs continuous action chunks at high frequency and is trained to produce smooth, reactive behaviour rather than discrete action tokens. A companion 'ER' (Embodied Reasoning) variant exposes intermediate reasoning, point/box predictions, and 3D grounding to drive planners.

Παράμετροι

Undisclosed (Gemini 2.0-class backbone with action head)

Δυνατότητες

• Generalist VLA built on Gemini 2.0 multimodal backbone
• Bimanual dexterous manipulation (Aloha 2, humanoids)
• Zero-shot generalisation to new objects and instructions
• Reactive closed-loop control with low latency
• Natural-language instruction following with chain-of-thought
• Multi-embodiment: arms, humanoids, mobile bases
• Tight integration with Gemini Robotics-ER for planning and grounding
• Public demos: folding origami, packing lunch boxes, slam-dunking miniature balls
• Best for: research collaborations on dexterous manipulation.

Εκπαίδευση & άδεια

Multi-embodiment robot demonstration corpus combining Google's internal datasets (RT-1/RT-2 lineage, Aloha 2 bimanual), partner humanoid data (Apptronik Apollo, Agility Digit) and Open-X-Embodiment style cross-embodiment teleoperation. Inherits Gemini 2.0's multimodal pretraining (text, image, video, code) at web scale.

Άδεια: Research-only - not publicly released; access via Google DeepMind research collaborations with selected hardware partners.

Δοκιμές ασφάλειας: Google DeepMind applies its standard Responsible AI / Frontier Safety Framework to Gemini Robotics, including pre-release safety evaluations, restricted access, and red-team probing for physical-harm and misuse scenarios. The Gemini Robotics technical report discusses safety guardrails and embodied red-teaming.