This is my implementation of the model behind RT-2. RT-2 leverages PALM-E as the backbone with a Vision encoder and language backbone where images are embedded and concatenated in the same space as the language embeddings. This architecture is quite easy to architect but suffers from a lack of deep understanding of both the unified multi modal representation or the individual modality representations.
RT-2 can be easily installed using pip:
pip install rt2The RT2 class is a PyTorch module that integrates the PALM-E model into the RT-2 class. Here are some examples of how to use it:
First, you need to initialize the RT2 class. You can do this by providing the necessary parameters to the constructor:
import torch
from rt2.model import RT2
# img: (batch_size, 3, 256, 256)
# caption: (batch_size, 1024)
img = torch.randn(1, 3, 256, 256)
caption = torch.randint(0, 20000, (1, 1024))
# model: RT2
model = RT2()
# Run model on img and caption
output = model(img, caption)
print(output) # (1, 1024, 20000)
RT-2 stands at the intersection of vision, language, and action, delivering unmatched capabilities and significant benefits for the world of robotics.
- Leveraging web-scale datasets and firsthand robotic data, RT-2 provides exceptional performance in understanding and translating visual and semantic cues into robotic control actions.
- RT-2's architecture is based on well-established models, offering a high chance of success in diverse applications.
- With clear installation instructions and well-documented examples, you can integrate RT-2 into your systems quickly.
- RT-2 simplifies the complexities of multi-domaster understanding, reducing the burden on your data processing and action prediction pipeline.
RT-2 integrates a high-capacity Vision-Language model (VLM), initially pre-trained on web-scale data, with robotics data from RT-2. The VLM uses images as input to generate a sequence of tokens representing natural language text. To adapt this for robotic control, RT-2 outputs actions represented as tokens in the model’s output.
RT-2 is fine-tuned using both web and robotics data. The resultant model interprets robot camera images and predicts direct actions for the robot to execute. In essence, it converts visual and language patterns into action-oriented instructions, a remarkable feat in the field of robotic control.
Datasets used in the paper
| Dataset | Description | Source | Percentage in Training Mixture (RT-2-PaLI-X) | Percentage in Training Mixture (RT-2-PaLM-E) |
|---|---|---|---|---|
| WebLI | Around 10B image-text pairs across 109 languages, filtered to the top 10% scoring cross-modal similarity examples to give 1B training examples. | Chen et al. (2023b), Driess et al. (2023) | N/A | N/A |
| Episodic WebLI | Not used in co-fine-tuning RT-2-PaLI-X. | Chen et al. (2023a) | N/A | N/A |
| Robotics Dataset | Demonstration episodes collected with a mobile manipulation robot. Each demonstration is annotated with a natural language instruction from one of seven skills. | Brohan et al. (2022) | 50% | 66% |
| Language-Table | Used for training on several prediction tasks. | Lynch et al. (2022) | N/A | N/A |
The unique capabilities of RT-2 open up numerous commercial applications:
- Automated Factories: RT-2 can significantly enhance automation in factories by understanding and responding to complex visual and language cues.
- Healthcare: In robotic surgeries or patient care, RT-2 can assist in understanding and performing tasks based on both visual and verbal instructions.
- Smart Homes: Integration of RT-2 in smart home systems can lead to improved automation, understanding homeowner instructions in a much more nuanced manner.
Contributions to RT-2 are always welcome! Feel free to open an issue or pull request on the GitHub repository.
For any queries or issues, kindly open a GitHub issue or get in touch with kyegomez.
@inproceedings{RT-2,2023,
title={},
author={Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski,
Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu,
Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alexander Herzog,
Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang,
Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch,
Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi,
Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong,
Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu,
and Brianna Zitkovich},
year={2024}
}RT-2 is provided under the MIT License. See the LICENSE file for details.