In the real world, the strong episode resetting mechanisms that are needed to train agents in simulation are unavailable. The resetting assumption limits the potential of reinforcement learning in the real world, as providing resets to an agent usually requires the creation of additional handcrafted mechanisms or human interventions. Recent work aims to train agents (forward) with learned resets by constructing a second (backward) agent that returns the forward agent to the initial state. We find that the termination and timing of the transitions between these two agents are crucial for algorithm success. With this in mind, we create a new algorithm, Reset Free RL with Intelligently Switching Controller (RISC) which intelligently switches between the two agents based on the agent’s confidence in achieving its current goal. Our new method achieves state-of-the-art performance on several challenging environments for reset-free RL.
To install all the necessary dependencies, inside of a virtual environment (Python 3.10), first install the EARL benchmark:
git clone https://github.com/dapatil211/earl_benchmark.git
cd earl_benchmark
pip install -e .
Next, install the rest of the dependencies in requirements.txt.
cd ..
pip install -r requirements.txt
The configs for running the RISC agents described in the paper can be found in the
folder risc/configs/. To launch an agent, go to the risc/ directory and run:
cd risc/
python main.py -c configs/risc_sd.yaml
We provide RISC configs for Sawyer Door, Sawyer Peg, Tabletop Manipulation, Minitaur, and Minigrid.
The current codebase logs to Wandb. We provide the raw numbers for our runs as well as the code to convert them into the plots seen in the paper. To create all the plots, simply run:
cd visualization
python create_figures.py
To create just one of the figures, run:
cd visualization
python create_figures.py -c <"figure_3_top" | "figure_3_bottom" | "figure_4_left" | "figure_4_right" | "figure_5" | "figure_6">
If you found our work useful, please cite our paper.
@inproceedings{
patil2024intelligent,
title={Intelligent Switching for Reset-Free {RL}},
author={Darshan Patil and Janarthanan Rajendran and Glen Berseth and Sarath Chandar},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=Nq45xeghcL}
}