We study how vision-language models trained on Internet-scale data can be incorporated directly into end-to-end robotic control to boost generalization and enable emergent semantic reasoning. Our goal is to enable a single end-to-end trained model to both learn to map robot observations to actions and enjoy the benefits of large-scale pretraining on language and vision-language data from the web. To this end, we propose to co-fine-tune state-of-the-art vision-language models on both robotic trajectory data and Internet-scale vision-language tasks, such as visual question answering. In contrast to other approaches, we propose a simple, general recipe to achieve this goal: in order to fit both natural language responses and robotic actions into the same format, we express the actions as text tokens and incorporate them directly into the training set of the model in the same way as natural language tokens. We refer to such category of models as vision-language-action models (VLA) and instantiate an example of such a model, which we call RT-2. Our extensive evaluation (6k evaluation trials) shows that our approach leads to performant robotic policies and enables RT-2 to obtain a range of emergent capabilities from Internet-scale training. This includes significantly improved generalization to novel objects, the ability to interpret commands not present in the robot training data (such as placing an object onto a particular number or icon), and the ability to perform rudimentary reasoning in response to user commands (such as picking up the smallest or largest object, or the one closest to another object). We further show that incorporating chain of thought reasoning allows RT-2 to perform multi-stage semantic reasoning, for example figuring out which object to pick up for use as an improvised hammer (a rock), or which type of drink is best suited for someone who is tired (an energy drink).
The paper investigates the integration of large-scale pretrained vision-language models (VLMs) into robotic control frameworks.
The study introduces a training method that includes robotic actions as tokenized input data, allowing robots to interpret and execute textual action sequences.
The RT-2 model showing improved generalization and reasoning capabilities is extensively tested and demonstrates superior performance compared to similar models.
RT-2 faces challenges such as dependency on the variety of skills in training data and the computational demands of large models.
Future work could involve strategies to reduce model size and increase generalizability by incorporating diverse training data, including human demonstrations.
High-capacity models pretrained on vast web-scale datasets have exhibited effective and powerful downstream task performance. Incorporating such models directly into end-to-end robotic control has the potential to vastly improve generalization and enable emergent semantic reasoning. This study probes the extent to which large-scale pretrained vision-language models (VLMs) can be integrated into robotic control, proposing a framework to train vision-language-action (VLA) models capable of interpreting and acting upon both linguistic and visual cues in the context of robotics.
The core hypothesis tested is the practicality of co-fine-tuning state-of-the-art VLMs on robotic trajectory data alongside Internet-scale vision-language tasks. This entails a training paradigm where robotic actions are tokenized and added to the training set, treated comparably to natural language tokens. The VLA model can then output textual action sequences in response to prompts, which the robot interprets and enacts. The RT-2 model, an instantiation of such VLA architecture, is then rigorously evaluated across thousands of trials to validate its performance and emergent capabilities, such as improved generalization, command interpretation, and basic semantic reasoning.
The investigation is situated within a landscape of recent works attempting to blend Language Models (LLMs) and VLMs into robotics. While these methods generally excel at high-level planning, they often fail to leverage the rich semantic knowledge during training for lower-level tasks. RT-2's approach is contrasted with those alternatives through its innovative real-time inference protocol and training strategy. Remarkably, RT-2 models demonstrate preferable generalization to novel objects, environments, and semantically varied instructions, significantly outperforming similar robotics-focused models.
Despite RT-2's promising abilities, the approach retains limitations such as its reliance on the range of physical skills present in the training robot data and the computational costs associated with running large models. Future research directions include potential solutions such as model quantization and distillation to improve efficiency. Moreover, greater availability of open-sourced VLMs and the ability to train on a broader array of data, including human demonstrations, could extend RT-2's generalizability and skillset further.
In conclusion, RT-2 suggests a powerful avenue through which robust pretraining on language and vision can enhance robotic control. By effectively transferring web-scale pretraining to robotic manipulation, RT-2 sets a benchmark for future vision-language-action models in the robotics domain.