TSRL (https://arxiv.org/abs/2306.04220) introduces a new offline reinforcement learning (RL) algorithm that leverages the fundamental symmetry of system dynamics to enhance performance under small datasets. The proposed Time-reversal symmetry (T-symmetry) enforced Dynamics Model (TDM) establishes consistency between forward and reverse latent dynamics, providing well-behaved representations for small datasets. TSRL achieves impressive performance on small benchmark datasets with as few as 1% of the original samples, outperforming recent offline RL algorithms in terms of data efficiency and generalizability.

To install the dependencies, use

    pip install -r requirements.txt

You can download the small samples from the ''utils/small_samples/'' directly.

Or if you want to generate by yourself:

    bash utils/generate_loco.sh # For the locomotion tasks

and

    bash utils/generate_adroit.sh # For the adroit tasks

You can train TDM simply from:

    bash TDM/train_loco.sh # For the locomotion tasks 

and

    bash TDM/train_adroit.sh #  For the adroit tasks

After you have your own small samples as well as a trained TDM model, you can run TSRL on D4RL tasks by:

    bash tsrl_loco.sh # For the locomotion tasks 

and

    bash tsrl_adroit.sh # For the locomotion tasks 

You can resort to wandb to login your personal account via export your own wandb api key.

export WANDB_API_KEY=YOUR_WANDB_API_KEY

and run

wandb online

to turn on the online syncronization.