DI-engine is a comprehensive reinforcement learning framework from OpenDILab that covers nearly every family of decision-making algorithms in one codebase. It provides the environments, policies, models and middleware needed to research, train and benchmark RL agents on PyTorch and JAX.
DI-engine is open source on GitHub with about 3,650 stars and releases continuing into 2025. Its Python-first, asynchronous design modularly integrates core concepts such as Env, Policy and Model, and it ships with thoroughly tested implementations across the RL landscape: classic deep RL like DQN, PPO and SAC; multi-agent methods like QMIX and MAPPO; imitation learning such as GAIL; offline RL including CQL and Decision Transformer; model-based RL up to DreamerV3; and exploration methods like RND and HER. It also includes LLM-plus-RL algorithms such as DPO, along with distributed training support and environments ranging from Atari and MuJoCo to SMAC and MiniGrid. The project is maintained by the OpenDILab community at Shanghai AI Laboratory and is used as the reference codebase behind several of the lab’s decision-intelligence research efforts, which keeps the implementations aligned with published results. Benchmark results for the major algorithms are published with the documentation, so users can sanity-check their own runs.
DI-engine is free and open source under the Apache-2.0 license. There are no paid tiers; you only supply your own compute for training.
Reinforcement learning researchers and engineers who want a single, well-organized framework to compare algorithms fairly and scale from quick prototypes to serious training runs.