The Official Code of "Deep Demonstration Tracing: Learning Generalizable Imitator Policy for Runtime Imitation from a Single Demonstration" (ICML'24). Visit our project page OSIL for more information.
Step1: We offer code for VPAM and Metaworld environment. You can download demo dataset from Google Drive to .data
Step2: install related packages
conda create -n ddt python=3.10
pip install -r requirements.txt
install environments
cd envs/gym_continuous_maze
pip install -e .
install experiments management tool RLAssitant
git clone https://github.com/polixir/RLAssistant.git
cd RLAssistant
pip install -e .
We validate the performance of DDT in both VPAM and Metaworld, which shows significantly better performance than the baselines with unforseen obstacle or without obstacle. Part of the results are as follows:
You can run DDT with the following command. For maze environment, the default config can be found in "configs/maze_mt.yaml". For Metaworld environment, the default config can be found in "configs/metaworld_mt.yaml".
- Maze:
python main_ddt.py --device cuda:0 --benchmark maze- Metaworld:
python main_ddt.py --device cuda:0 --benchmark metaworld