RT-2

Translates free-form natural language instructions into executable robot control signals by processing robot camera observations alongside text commands through a unified vision-language-action transformer. The model encodes robot actions as text tokens within the language modeling framework, enabling the same transformer architecture to handle both semantic understanding and motor control generation. This co-fine-tuning approach preserves pre-trained vision-language knowledge while adding robotic trajectory supervision, allowing the model to ground language semantics directly to physical actions.

Leverages pre-trained vision-language model knowledge to recognize and manipulate objects not present in the robot training dataset by grounding language descriptions to visual features learned from internet-scale data. When given an instruction like 'pick up the extinct animal,' the model maps the semantic concept to visual features of novel objects through language understanding rather than explicit object-specific training. This capability emerges from co-fine-tuning robotic trajectories with vision-language tasks, allowing the model to apply learned semantic relationships to new physical scenarios.

Performs relative comparisons and superlative reasoning on objects in the robot's visual field by leveraging language model understanding of comparative semantics. The model can interpret instructions like 'pick up the smallest object' or 'place it closest to the red cube' by reasoning about spatial and attribute relationships between multiple objects in a single image. This capability combines vision-language understanding with robotic action generation, allowing the model to compute relative properties and select appropriate targets without explicit comparative logic programming.

Generates intermediate reasoning steps before producing final robot actions, enabling decomposition of complex tasks into semantic sub-goals. When processing instructions like 'use an improvised tool to reach the object,' the model can emit chain-of-thought tokens that reason about available tools, their properties, and applicability before selecting and executing an action. This approach leverages the language model's ability to generate text reasoning steps, then grounds those steps in robotic actions, allowing the model to handle multi-stage semantic reasoning without explicit task decomposition modules.

Combines robotic trajectory data with internet-scale vision-language tasks during training while preserving the pre-trained vision-language model's learned representations. Rather than replacing the original model with robot-specific weights, co-fine-tuning maintains the vision and text encoder knowledge while adding robotic action supervision, allowing the model to retain semantic understanding from web-scale data while learning action grounding. This hybrid training approach encodes actions as text tokens to fit into the standard language modeling framework, enabling efficient knowledge transfer from vision-language pretraining to robotic control.

Encodes robot actions as discrete text tokens within the language model's vocabulary, enabling actions to be generated using the same transformer decoder as natural language. Rather than predicting continuous control values or using separate action heads, the model maps each possible robot action to a unique token, allowing the language modeling framework to handle both semantic understanding and action generation. This unified representation simplifies the architecture and enables joint training on language and robotic tasks without specialized control modules.

Grounds abstract semantic concepts from vision-language models to concrete physical robot actions by training on paired robot observations and action trajectories. The model learns to map visual features and language semantics (learned from internet-scale data) to specific motor commands, creating a bridge between high-level semantic understanding and low-level robot control. This grounding process occurs during co-fine-tuning, where robotic trajectory supervision teaches the vision-language model which actions correspond to which visual and linguistic inputs.

Provides evaluation infrastructure for assessing robot control models across 6,000 diverse trials covering different objects, instructions, and scenarios. This evaluation framework enables systematic assessment of generalization, semantic understanding, and action accuracy across a large test set. The scale of evaluation (6,000 trials) suggests comprehensive coverage of task variations, though specific metrics, success criteria, and baseline comparisons are not disclosed in available documentation.

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.