conda env create -f environment.yaml
CuGRO tested on two classical benchmarks: MuJoCo and Meta-World. The collected dataset can be downloaded here
Running experiments based on our code can be quite easy. You can run all benchmarks by executing the shell file:
sh run.sh
Or you can execute the following command to run CuGRO, take "cheetah_vel" as an example :
-
Train the state generator and the behavior generator.
CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 main-gene.py --env "cheetah_vel" --data_mode "gene" --actor_type "large" --diffusion_steps 100
-
Train the critic model and plot the results in the logs of all sequential tasks:
python critic.py --env "cheetah_vel" --data_mode "gene" --actor_type "large" --diffusion_steps 100 --gpu 0