PyTorch re-implementation of RT-1. This model is based on 2022 paper "RT-1: Robotics Transformer for Real-World Control at Scale". This implementation mostly follows the original codes but I made some changes for PyTorch compatibility or simplification.
- Film efficient net based image tokenizer backbone
- Token learner based compression of input tokens
- Transformer for end to end robotic control
- Testing utilities
- Translated TensorFlow implementation into PyTorch implementation.
- Used spaces of OpenAI gym to define observation and action variables instead of using tensor_spec of TF-Agents.
- Abolished 4-d constraint of some specs in transformer_network.py
- Added useful functions and put those in utils.py.
- Added extra comments.
- Didn't implement the add_summaries function of transformer_network.py which is for tensorboard logging.
- Omitted some variables, functions, classes, and tests including those that are no longer necessary for this repository.
I didn't implement squence_agent.py and its related codes. Because those codes are for training and inference that use the tf_agent library, so I figure it is not necessary in PyTorch implementation. This repository enables you to use RT-1 just like the other Reinforcement learning codes that utilize PyTorch and OpenAI Gym, without the tf_agent library.
This library is licensed under the terms of the Apache license.