RT-2

RT-2 maps robot observations (images) and natural language commands directly to executable robot actions by leveraging a transformer-based vision-language-action architecture that co-trains on Internet-scale vision-language data alongside robot trajectory data. Actions are represented as discrete text tokens integrated into the language model's vocabulary, enabling the model to reason about visual scenes and language semantically before outputting action sequences. This approach transfers web-scale knowledge (VQA, visual reasoning) to robotic control without requiring explicit action space engineering.

RT-2 generalizes to natural language commands not present in its robot training data by applying semantic reasoning learned from Internet-scale vision-language tasks. The model interprets novel command phrasings (e.g., 'place object on the icon' or 'on the number 5') by decomposing them into visual and semantic concepts it has learned from VQA and general vision-language co-training, then mapping those concepts to appropriate robot actions. This capability emerges from the co-training approach rather than explicit command parsing or semantic slot-filling.

RT-2 performs chain-of-thought reasoning over visual observations and natural language instructions to decompose complex manipulation tasks into sub-goals and select appropriate actions. For example, when instructed to 'use an improvised hammer to break something,' the model reasons about which object could serve as a hammer, how to grasp it, and how to apply it — this reasoning emerges from the transformer's ability to process visual and linguistic context jointly. The text-token action representation allows the model to express intermediate reasoning steps as part of the action sequence.

RT-2 selects objects based on comparative properties (smallest, largest, closest to another object, matching a description) by reasoning about visual relationships and semantic attributes. The model processes the visual scene, understands the comparative property being requested, and identifies the target object — this capability emerges from vision-language pre-training on tasks like VQA that require comparative reasoning. The selected object is then grounded to robot actions for manipulation.

RT-2 adapts robot behavior based on contextual information inferred from visual observations and task descriptions. For example, when instructed to 'select an appropriate drink for a sleepy person,' the model reasons about the person's state, the available drinks, and task-specific appropriateness — this contextual reasoning emerges from the vision-language pre-training's ability to understand human states, object properties, and task semantics. The model then selects and manipulates the appropriate object.

RT-2 generalizes to object categories not seen during robot training by leveraging semantic understanding from Internet-scale vision-language pre-training. When encountering a novel object, the model recognizes its visual features and semantic properties (learned from web-scale data), maps those properties to appropriate manipulation strategies, and executes actions — this transfer occurs without explicit fine-tuning on the novel object category. The co-training approach ensures that visual and semantic knowledge from web-scale data directly informs robot action selection.

RT-2 is trained through a co-training approach that jointly optimizes on Internet-scale vision-language tasks (VQA, visual reasoning) and robot trajectory data, maintaining some original vision-language data during training. This approach transfers semantic and visual understanding from web-scale data to robotic control by representing actions as text tokens integrated into the language model vocabulary. The co-training ensures that the model learns generalizable visual and semantic concepts before specializing to robot-specific action prediction.

RT-2 represents robot actions as discrete text tokens integrated into the language model's vocabulary, enabling the model to predict actions using the same token prediction mechanism as language generation. This approach allows actions to be expressed alongside natural language reasoning and intermediate steps, and leverages the transformer's language modeling capabilities for action prediction. Actions are decoded from text tokens into robot-specific motor commands through an integration layer.

RT-2 grounds natural language instructions to specific visual elements in robot observations by jointly processing images and text through the vision-language transformer. When given an instruction like 'pick up the red cube,' the model identifies the red cube in the visual scene and predicts actions to manipulate it — this grounding emerges from the transformer's ability to attend to relevant visual regions while processing language. The model learns to align language tokens with visual features through co-training on vision-language tasks.

RT-2 was evaluated on 6,000+ robotic manipulation trials to assess performance on object picking, generalization to novel objects, out-of-distribution command interpretation, and comparative reasoning tasks. The evaluation protocol tests the model's ability to follow natural language instructions in real robotic scenarios, though specific quantitative metrics, success rates, and comparison to baselines are not publicly documented. The evaluation scale demonstrates the feasibility of the approach but lacks detailed performance characterization.