Google RT-2-X

Sobre Google DeepMind

Fundado em 2010 · London, UK / Mountain View, USA

RT-2-X is Google DeepMind's flagship Robotic Transformer 2 (RT-2) model retrained on the Open-X-Embodiment dataset - the first large-scale, multi-institution effort to assemble a unified robot-learning dataset spanning many labs and robots. Open-X-Embodiment was organised in 2023 by Google DeepMind together with 21+ academic and industry institutions (Stanford, UC Berkeley, CMU, MIT, Toyota Research Institute, etc.), producing the RT-X dataset of ~1 million trajectories across 22 robot embodiments. RT-2-X extends RT-2's Vision-Language-Action recipe - using a PaLM-E / PaLI-X style VLM as backbone and emitting actions as text tokens - to this cross-embodiment corpus, demonstrating positive transfer across robots and a new state of the art on generalist manipulation at the time of release. RT-2-X is research-only and not publicly callable; it remains a key academic reference and the conceptual parent of subsequent open VLAs.

Arquitetura

Vision-Language-Action transformer (PaLI / PaLM-E backbone, discrete action tokens)

RT-2-X follows the RT-2 design: a large Vision-Language Model (PaLI-X or PaLM-E) is co-fine-tuned on web-scale vision-language data and on robot demonstration data, where robot actions are tokenised as strings of natural-language-like tokens (each action dimension binned and rendered as a token). The same next-token prediction objective therefore trains the model on both internet-scale image-text data and on robot trajectories, allowing the resulting policy to inherit web knowledge (object semantics, OCR, common sense) and route it to motor commands. RT-2-X is the version of this recipe trained on the Open-X-Embodiment / RT-X dataset - ~1 million trajectories across 22 robot embodiments - rather than only on Google's internal kitchen-robot dataset. Public results report 5B and 55B variants, with the 55B model showing the strongest generalisation, especially when prompted with unseen language commands or unseen object combinations.

Parâmetros

Up to 55B (RT-2-X variants: 5B and 55B)

Capacidades

• Generalist VLA trained on Open-X-Embodiment (22 robots)
• Inherits web-scale knowledge from PaLI / PaLM-E backbones
• Discrete action-token decoding (text-like vocabulary)
• Positive transfer across robot embodiments
• Strong emergent semantic reasoning (e.g. 'pick up the extinct animal')
• 5B and 55B parameter variants
• Reference architecture for the modern VLA paradigm
• Co-training on internet data + robot demos
• Best for: research, citation, conceptual baseline for VLAs.

Treinamento & licença

Co-trained on internet-scale vision-language data (PaLI / PaLM-E corpora) plus ~1 million robot trajectories from the Open-X-Embodiment (RT-X) dataset across 22 robot embodiments. Action targets are tokenised continuous controls.

Licença: Research-only - Google DeepMind has not publicly released the RT-2-X weights, code or API. Some Open-X-Embodiment data and smaller RT-X reproductions are available, but the proprietary RT-2-X checkpoints are not.

Testes de segurança: Covered under Google's standard Responsible AI process and frontier-safety evaluations. Robot demos were filmed in supervised lab settings with force-limited hardware; no public RSP-style document specific to RT-2-X.