SpringGrasp

Website Paper

Synthesizing Compliant, Dexterous Grasp under Shape Uncertainty

Optimization based compliant grasp synthesis using only single depth image.

Installation

Install basic python dependencies

pip install -r requirements.txt

Install thirdparty tools

cd thirdparty
cd differentiable-robot-model
pip install -e .
cd TorchSDF
bash install.sh

Install curobo for arm motion planning and collision avoidance [optional]

mkdir curobo_ws

Download and unzip customized curobo inside curobo_ws

Follow install instruction of each package in their README.md.

File structure

root directory
  ├── assets  
  │   └── // folders for real scanned object
  ├── data  
  │   └── // folders for data after optimization
  ├── gpis_states  
  │   └── // state data for restoring and visualizing gaussian process implicit surface
  ├── thirdparty  
  │   ├── // dependent third-party package
  |   └── TorchSDF
  |      └── // Compute SDF of mesh in pytorch
  |   └── differentiable-robot-model
  |      └── // Differentiable forward kinematics model
  ├── [curobo_ws] // customized curobo motion planner
  │   ├── curobo
  |   └── nvblox
  |   └── nvblox_torch
  ├── gpis
  |   ├── 3dplot.py // Visualizing GPIS intersection and its uncertainty
  |   └── gpis.py // Definition for Gaussian process implicit surface
  ├── spring_grasp_planner // Core implementation
  |   ├── initial_guesses.py // Initial wrist poses
  |   ├── metric.py // Implementation of spring grasp metric
  |   └── gpis.py // Different optimizers for spring grasp planner
  ├── utils
  |   └── // Ultilities to support the project
  ├── process_pcd.py // Processing pointclouds from different cameras
  ├── optimize_pregrasp.py // Running compliant pregrasp optimization
  └── verify_grasp_robot.py  // verifying pregrasp on hardware, kuka iiwa14 + left allegro hand.

Usage

Plan grasp for pre-scanned objects

Process pointcloud

python process_pcd.py --exp_name <obj_name>

Plan and visualizing grasp

python optimize_pregrasp.py --exp_name <obj_name>

Customized object

Plan and visualize grasp

python optimize_pregrasp.py --exp_name <obj_name> --pcd_file <path to your ply pointcloud>

Plan reaching motion for Kuka arm (require curobo installation and scene configuration)

python traj_gen.py --exp_name <obj_name>

Execute grasp on real robot(require support for kuka iiwa14 + left allegro hand)

python verify_grasp_robot.py --exp_name <obj_name>

Using customized scene and deploy on hardware

How to deploy on hardware varies case by case, if you need help with using Kuka iiwa14 + allegro hand or run into troubles with coordinate system convention please contact: ericcsr [at] stanford [dot] edu

Citation

If you find our code useful, please cite the following

@misc{chen2024springgrasp,
      title={SpringGrasp: An optimization pipeline for robust and compliant dexterous pre-grasp synthesis}, 
      author={Sirui Chen and Jeannette Bohg and C. Karen Liu},
      year={2024},
      eprint={2404.13532},
      archivePrefix={arXiv},
      primaryClass={cs.RO}
}

Stanford-TML/SpringGrasp_release