VGCP: learning Inverse Dynamics Model using Inverse Reinforcement Learning
Preparation
1. Encoder for reward computation: The encoder is pretrained using TCC, and each is for one task. Download pretrain_runs.zip
2. Datasets for conditioning goals: The goals in observation are from a randomly sampled sequence in demonstrations. They are psuedo-goals. Download and put datasets under VGCP.
3. Finally, prepare video generative models. ckpt 38 is trained on hammer, assembly, basketball, but not all tasks. It is tested to show good quality on the trained tasks. ckpt 35 and 36 are trained on all tasks, but not tested w.r.t. quality. Put video_model_ckpts holding ckpts under VGCP.
cd VGCP, run "python rl_mw_learned_reward.py --pretrained_path='/home/lei/Documents/game/multi-task-tcc/experiments/pretrain_runs/dataset=metaworld_algo=xirl_uid=assembly' --env_name=assembly". Some paths may need to be updated to the machine of your own. Replace assembly with other task names.
