rlbench_exp.md

🚶Evaluate Proprioceptive Pointcloud Transformer (PPT) only on RLBench

First you should install RLBench dependencies following this repository.

Generate demonstrations

1. Change the environments in gensim2/env/utils/rlbench.py
2. Run python scripts/rlbench_data_collection_mp.py --nprocs 32 --obs_mode pointcloud --random --asset_id 'random' ---save --num_episodes 100 (use --save to save collected data onto disks, --random is for randomization)

Train&Test multi-task policies

1. Choose the config in gensim2/agent/experiments/configs/env and replace the name in gensim2/agent/experiments/configs/config.yaml, Line 14. Parameter domains determine the name of dataset.
2. Run

python agent/run.py  \  # Train
  suffix=name_of_the_run \  
  dataset.action_horizon=4 \ # predicted action sequence, has to be 1 for MLP head
  dataset.observation_horizon=3 \ # historical observation horizon
  train.total_epochs=0 # training epochs

python agent/run.py  \  # Test
  suffix=name_of_the_run \  
  dataset.observation_horizon=3 \ 
  train.total_epochs=0 \ # to prevent from re-training
  train.pretrained_dir=relative_path_to_the_model.pth \
  network.openloop_steps=${openloop_steps} \  # openloop actions
  temporal_agg=False; # trick as in ACT, this will make openloop_steps=1 only