Grounded Policy Network

A replicated version of GR1 with added mask input.

Setup

Please follow the main setup instruction to prepare RoboGround's Conda environment.

conda activate roboground

# install the additional dependency
pip install git+https://github.com/openai/CLIP.git

Training & Evaluation

Training the model takes approximately 7 days using 8 NVIDIA RTX 4090 GPUs.

Simply start training with: bash scripts/train.sh. And then run the evaluation using: bash scripts/eval.sh.