DarkDawn233/D2R_MTRL

Codes of D2R accompanying the paper "Not All Tasks Are Equally Difficult: Multi-Task Deep Reinforcement Learning with Dynamic Depth Routing"(AAAI 2024).

Code Instruction for D2R

This instruction hosts the PyTorch implementation of "Not All Tasks Are Equally Difficult: Multi-Task Deep Reinforcement Learning with Dynamic Depth Routing" (AAAI 2024, Link) with the Meta-World benchmark.

NOTE: Since MetaWorld is under active development, we perform all the experiments on the following commit-id: https://github.com/Farama-Foundation/Metaworld/commit/04be337a12305e393c0caf0cbf5ec7755c7c8feb

Setup

1. Set up the working environment:

Required packages: pytorch==1.13.1, json5, tensorboardX, posix_ipc, scipy

pip install -r requirements.txt

2. Set up the Meta-World benchmark:

Please follow the instructions to install the mujoco-py package first.

pip install git+https://github.com/Farama-Foundation/Metaworld.git@04be337a12305e393c0caf0cbf5ec7755c7c8feb

Training

To train D2R on the MT10-Rand setting,

python train.py --config config/mt10/depthroute_rand.json --id mt10_rand --seed 1

Change the config accordingly for other setting (e.g. MT50-Rand).

All results will be saved in the log folder.

The config file config/mt10/depthroute_rand.json contains default hyperparameters for D2R.

See Also

See Meta-World, mujoco-py for additional instructions.