Code release for Efficient Planning in a Compact Latent Action Space (ICLR2023). We recommend to check out the porject page for a quick summary of the method and results.
Abstract:
Planning-based reinforcement learning has shown strong performance in tasks in discrete and low-dimensional continuous action spaces. However, planning usually brings significant computational overhead for decision-making, and scaling such methods to high-dimensional action spaces remains challenging. To advance efficient planning for high-dimensional continuous control, we propose Trajectory Autoencoding Planner (TAP), which learns low-dimensional latent action codes with a state-conditional VQ-VAE. The decoder of the VQ-VAE thus serves as a novel dynamics model that takes latent actions and current state as input and reconstructs long-horizon trajectories. During inference time, given a starting state, TAP searches over discrete latent actions to find trajectories that have both high probability under the training distribution and high predicted cumulative reward. Empirical evaluation in the offline RL setting demonstrates low decision latency which is indifferent to the growing raw action dimensionality. For Adroit robotic hand manipulation tasks with high-dimensional continuous action space, TAP surpasses existing model-based methods by a large margin and also beats strong model-free actor-critic baselines.
All python dependencies are in environment.yml. Install with:
conda env create -f environment.yml
conda activate tap
pip install -e .
Please note that installing mujoco-py sometimes requires a bit more efforts. See this page for more details.
There are three scripts (loco.sh, adroit.sh and antmaze.sh) to run experiments for locomotion control, antmaze and adroit. The selected task will be run 5 times with different training seeds and each trained model will then be evaluated with 20 test seeds (20 episodes).
To break it down:
- train the encoder and decoder:
python scripts/train.py --dataset $data --exp_name $name-$round --tag development --seed $round
- train the auto-regressive prior:
python scripts/trainprior.py --dataset $data --exp_name $name-$round
- evaluate the trained models:
for i in {1..20};
do
python scripts/plan.py --test_planner beam_prior --dataset $data --exp_name $name-$round --suffix $i --beam_width 64 --n_expand 4
done
- report the results:
python plotting/read_results.py --exp_name $name --dataset $data
GLIBCXX_2.4.29 not being found
Solution:
ln -s /usr/lib/x86_64-linux-gnu/libstdc++.so.6 libstdc++.so.6
conda install -c conda-forge glew sudo apt-get install libglew-dev
running build_ext
building 'mujoco_py.cymj' extension
...
Solution:
conda install -c conda-forge mesalib
The organization of this repo is based on trajectory-transformer repo. Some of the visualization is based on diffuser repo.
@article{jiang2023tap,
title={Efficient Planning in a Compact Latent Action Space},
author={Jiang, Zhengyao and Zhang, Tianjun and Janner, Michael and Li, Yueying and Rocktäschel, Tim and Grefenstette, Edward and Tian, Yuandong},
journal={ICLR2023},
year={2023},
}
