SUGARL

Code for NeurIPS 2023 paper Active Vision Reinforcement Learning with Limited Visual Observability, by Jinghuan Shang and Michael S. Ryoo.

We propose Sensorimotor Understanding Guided Active Reinforcement Learning (SUGARL) to solve ActiveVision-RL tasks. We also introduce Active-Gym, a convenient library that modifies existing RL environments for ActiveVision-RL, with Gymnasium-like interface.

[Paper] [Project Page] [Active-Gym]

Dependency

conda env create -f active_rl_env.yaml

We highlight Active-Gym developed by us to support Active-RL setting for many environments.

Usage

• General format:

  cd sugarl       # make sure you are under the root dir of this repo
  bash ./scripts/<any_setting.sh> agent/<any_agent_valid_for_that_setting.py>

• Reproduce our experiments:

  cd sugarl       # make sure you are under the root dir of this repo
  bash ./scripts/robosuite_series.sh agent/<any_agent_valid_for_that_setting.py>
  bash ./scripts/atari_series.sh agent/<any_agent_valid_for_that_setting.py>
  bash ./scripts/atari_series_5m.sh agent/<any_agent_valid_for_that_setting.py>
  bash ./scripts/atari_wp_series.sh agent/<any_agent_valid_for_that_setting.py>
  bash ./scripts/dmc_series.sh agent/<any_agent_valid_for_that_setting.py>

  For example, to run SUGARL-DQN on Atari

  bash ./scripts/atari_series.sh agent/dqn_atari_sugarl.py

• Sanity checks: they run through the whole process with only a tiny amount of training to check bugs

  cd sugarl       # make sure you are under the root dir of this repo
  bash ./scripts/atari_test.sh agent/<any_agent_valid_for_your_test.py>
  bash ./scripts/dmc_test.sh agent/<any_agent_valid_for_your_test.py>
  bash ./scripts/robosuite_test.sh agent/<any_agent_valid_for_your_test.py>

All experiment scripts automatically scale all tasks to your GPUs. Please modify the gpu behavior (CUDA_VISIBLE_DEVICES=<x>) in the script if

• you want to run jobs on certain GPUs
• either VRAM or RAM is not sufficient for scaling all jobs

In the provided scripts, 26 Atari games are in parallel, with sequentially executing each seed. 6 DMC environments x 5 seeds are all in parallel. Please do check the available RAM and VRAM on your machine before starting.

Notes

Naming:

All agents are under agent/, with the name format <base_algorithm>_<env>_<variant>.py. Each file is an individual entry for the whole process. We support DQN, SAC, and DrQ for base algorithms.

All experiment scripts are under scripts/, with the format <env>_<setting>.sh Please ensure that the env and setting match the agent when launching jobs.

Resource requirement reference (SUGARL):

• Atari: for each game with 100k replay buffer: ~18G RAM, <2G VRAM

• DMC: for each task with 100k replay buffer: ~18G RAM, <3G VRAM

• Robosuite: for each task with 100k replay buffer: ~54G RAM, 4.2G VRAM

Coding style:

We follow the coding style of clean-rl so that modifications on one agent would not affect others. This does introduce lots of redundency, but is so much easier for arranging experiments and evolving the algorithm.

Citation

Please consider cite us if you find this repo helpful.

@article{shang2023active,
    title={Active Reinforcement Learning under Limited Visual Observability},
    author={Jinghuan Shang and Michael S. Ryoo},
    journal={arXiv preprint},
    year={2023},
    eprint={2306.00975},
}

Acknowledgement

We thank the implementation of clean-rl.