Basically, this is decision transformers, but with the Mamba architecture. Since Mamba has the 'convolutional' mode for parallelizable training, it is a prime candidate for offline RL. Since it also has a 'recurrent' mode, it can do fast inference, which would be ideal for online training.

1. Train Mamba on offline Mujoco dataset [ ] (working on training code right now)
2. Evaluate Mamba against base Decision Transformer [ ]
3. Fine-tune Mamba using an online RL algorithm (probably PPO) [ ]

Contributions welcome

Install requirements.txt with pip install -r requirements.txt. I also ran these commands to installs these packages in Ubuntu:

sudo apt-get update
sudo apt-get install -y \
    libgl1-mesa-dev \
    libgl1-mesa-glx \
    libglew-dev \
    libosmesa6-dev \
    software-properties-common \
    patchelf \
    xvfb

Honestly I don't know what any of them do, but they were in the list of packages to install in the train your first decision transformer notebook so I installed them.

It's super easy to install the newer version of mujoco. Barely an inconvenience. But to access older (v3 and below) versions of the mujoco envs... You have to install mujoco-py and that is a pain. Here's what I found:

1. If you get an error trying to compile mujoco-py the first time it runs, try Cython < 3 (which I have in the requirements.txt file). See this issue for more info.
2. If you get an error like GLIBCXX_3.4.29 not found, read this StackOverflow post. Basically, delete libstdc++.so.6 (or whatever file is causing the error) from your anaconda_install_dir/envs/env_name/lib/ folder. If it is coming from a system path, not Anaconda, you probably shouldn't delete it.

Okay so the recurrent mode is slower than the normal mode for some reason, even though when the sizes are about the same for text generation, inference for mamba is about as fast as inference for DT and it is a lot faster when the graph is cached (like, 6x).