Official code of RA-L 2024 paper.

The total framework consists of 4 parts: Graspable Region Explorer, Grasp Detection, Grasp Encoder and RL Policy.

This code has been tested on Ubuntu20.04 with Cuda 11.7, Python3.9 and Pytorch 1.13.1. Environment: create a conda environment with necessary packages specified by requirements.txt. Run the following code to generate the seperate gap_rl package:

pip install -e .

Our framework combines a LocalGrasp model for fast and object-centric grasp detection. You can configure it under the localgrasp directory.

Visualization Demo: You can run the following code to visualize the trajectories ({'random2d', 'line', 'circular', 'bezier2d'} as {'Rotation', 'Line', 'Circular', 'Random'} in the paper):

cd gap_rl/examples
python test_env_trajs.py

Trajectories Generation: gap_rl/utils/traj_utils.py.

Graspable Region Explorer: gap_rl/sim2real/sim2real_traj_following.py -> track_grasps.

Grasp Detection: gap_rl/localgrasp/.

Grasp Encoder: gap_rl/algorithms/rl_utils.py -> CustomGraspPointGroupExtractor.

RL Policy： training -> gap_rl/algorithms/scripts/sac_train.py, evaluation -> gap_rl/algorithms/scripts/sac_LoG_dynamic_eval.py.

Sim2Real Implementation: gap_rl/sim2real/sim2real_traj_following.py.

In simulation, points filtering + grasp detection totally takes about 60-90ms per step, rendering and physical computation takes about 20-30ms per step, and thus totally it takes >100ms to collect one sample, which decrease the sample efficiency, considering that our training procedure need > 2M steps, which totally may take >48h for single training process, We think it's too long.

So We have tried another way to optimize the time burden -- pre-collect the grasps under different viewpoints and save them in the json files, just as the following picture shows. And during the training procedure, we take the pre-collected grasps from the nearest 4 viewpoints. The assumption here is that our LoG model detect grasps mainly depend on the viewpoints (distance is not important since the point cloud is normalized in the object points center.)

In this way, and also we utilize the Vectorized Environments from Stable Baselines 3, one single training process has been reduced from >48h to ~6h. Our tsne analysis and real experiments have demonstrated that it is useful and efficient in the dynamic settings.

Grasp Collection: gap_rl/examples/gen_LoG_grasps.py.

4-View Grasp Selection: gap_rl/envs/pick_single.py -> _compute_near_grasps.

If you find this code useful for your research, please use the following BibTeX entry.

@article{xie2024gap,
  title={GAP-RL: Grasps as Points for RL Towards Dynamic Object Grasping},
  author={Xie, Pengwei and Chen, Siang and Chen, Qianrun and Tang, Wei and Hu, Dingchang and Dai, Yixiang and Chen, Rui and Wang, Guijin},
  journal={IEEE Robotics and Automation Letters (RA-L)},
  year={2024}
}