Tools & Libraries

We provide open code implementations for most of our research, please check our papers for related codes. In addition, we aim to develop easy-to-use and comprehensive algorithm libraries and tools to accelerate the real-world deployment of advanced data-driven decision-making methods.

Data-Driven Control Lib (D2C) is a library for data-driven decision-making & control based on state-of-the-art offline reinforcement learning (RL), offline imitation learning (IL), and offline planning algorithms. It is a platform for solving various decision-making & control problems in real-world scenarios. D2C is designed to offer fast and convenient algorithm performance development and testing, as well as providing easy-to-use toolchains to accelerate the real-world deployment of SOTA data-driven decision-making methods.

The current supported offline RL/IL algorithms include (more to come):

• Twin Delayed DDPG with Behavior Cloning (TD3+BC)
• Distance-Sensitive Offline Reinforcement Learning (DOGE)
• Dynamics-Aware Hybrid Offline-and-Online Reinforcement Learning (H2O)
• Sparse Q-learning (SQL)
• Policy-guided Offline RL (POR)
• Offline Reinforcement Learning with Implicit Q-Learning (IQL)
• Discriminator-Guided Model-Based Offline Imitation Learning (DMIL)
• Behavior Cloning (BC)

Features:

• D2C includes a large collection of offline RL and IL algorithms: model-free and model-based offline RL/IL algorithms, as well as planning methods.
• D2C is highly modular and extensible. You can easily build custom algorithms and conduct experiments with it.
• D2C automates the development process in real-world control applications. It simplifies the steps of problem definition/mathematical formulation, policy training, policy evaluation and model deployment.

Library Information:

• The library is available in https://github.com/AIR-DI/D2C.
• The tutorials and API documentation are hosted on air-d2c.readthedocs.io.

OneRL: Event-driven fully distributed reinforcement learning framework proposed in “A Versatile and Efficient Reinforcement Learning Approach for Autonomous Driving” (https://arxiv.org/abs/2110.11573) that can facilitate highly efficient policy learning in RL-based tasks.

• Super fast RL training! (15~30min for MuJoCo & Atari on single machine)
• State-of-the-art performance
• Scheduled and pipelined sample collection
• Completely lock-free execution
• Fully distributed architecture
• Full profiling & overhead identification tools
• Online visualization & rendering
• Support multi-GPU parallel training
• Support exporting trained policy to ONNX for faster inference & deployment