Skip to content

Latest commit

 

History

History
76 lines (50 loc) · 2 KB

GETTING_STARTED_RLBENCH.md

File metadata and controls

76 lines (50 loc) · 2 KB
Raw

Getting Started with RLBench

There are three simulation setups in RLBench: 1) PerAct, 2) GNFactor, and 3) Hiveformer. GNFactor uses exactly the same setup as PerAct. Both have different succes conditions and 3D object models than Hiveformer.

Before training/testing on each setup, please install the RLBench library correspondingly.

Train and evaluate on RLBench with the Peract/GNFactor setup

Step 0: Prepare data on RLBench

See Preparing RLBench dataset

Step 1: Install RLBench with the PerAct setup

> git clone https://github.com/MohitShridhar/RLBench.git
> git checkout -b peract --track origin/peract
> pip install -r requirements.txt
> pip install -e .

Remember to modify the success condition of close_jar task in RLBench, as the original condition is incorrect. See this pull request for more detail.

Step 2: Train the policy

  • Train 3D Diffuser Actor with the PerAct setup
> bash scripts/train_keypose_peract.sh
  • Train 3D Diffuser Actor with the GNFactor setup
> bash scripts/train_keypose_gnfactor.sh

We also provide training scripts for Act3D.

  • Train Act3D with the PerAct setup
> bash scripts/train_act3d_peract.sh
  • Train Act3D with the GNFactor setup
> bash scripts/train_act3d_gnfactor.sh

Step 3: Test the policy

  • Test 3D Diffuser Actor with the PerAct setup
> bash online_evaluation_rlbench/eval_peract.sh
  • Test 3D Diffuser Actor with the GNFactor setup
> bash online_evaluation_rlbench/eval_gnfactor.sh

We also provide testing scripts for Act3D.

  • Test Act3D with the PerAct setup
> bash online_evaluation_rlbench/eval_act3d_peract.sh
  • Test Act3D with the GNFactor setup
> bash online_evaluation_rlbench/eval_act3d_gnfactor.sh