Skip to content

We present Visual Pressure Estimation for Robots (ViPER), a neural network capable of visually estimating contact pressure for a soft gripper in the wild using only an RGB image from an eye-in-hand camera. ViPER enables precision grasping of small objects (e.g., a paperclip) on unseen cluttered surfaces.

Notifications You must be signed in to change notification settings

Healthcare-Robotics/ViPER

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Visual Pressure Estimation for Robots (ViPER)

Paper | Video | Dataset/Models

ViPER is a neural network that visually estimates contact pressure given an RGB image from an eye-in-hand camera. ViPER enables precision grasping of small objects (e.g. a paperclip) on unseen cluttered surfaces.

alt text

Setup (Tested in Ubuntu 20.04)

  • Clone this repo, as well as this one, and place them in the same directory.
  • Install the requirements in requirements.txt.
  • Install remaining dependencies: pip install open3d trimesh shapely mapbox_earcut
  • Install the Sensel API from the deb file in this repo.
  • Download the dataset and models here.
  • Place folders in the dataset in /data (if training) and the folders in the models folder in /checkpoints.

Collecting data

There are 4 types of data to capture:

  • train: Images that are paired with both pressure ground truth and force/torque ground truth
  • test: Same as train, but for evaluating the network on predicting pressure
  • weak_train: Images that are paired with only force/torque ground truth
  • weak_test: Same as train, but for evaluating the network on predicting only force/torque

Running the data collection script

  • python -m recording.capture_data --config <yaml file name from /config> --view --robot_state --stage <train, test, weak_train, weak_test> --folder <pick a name>
  • see keyboard_teleop() in robot/robot_control.py for keyboard controls.

Training

  • The training script will save a folder containing models from the same run. The folder is named <config>_<index> and the models are named model_<epoch>.
  • Run the training script: python -m prediction.trainer_weak --config <yaml file name from /config>

Robot (Hello Robot Stretch)

To run on a Hello Robot Stretch, clone this repo on the robot. Then, verify the robot and PC are on the same network and that the IPs match those in /robot/zmq_client.py Run python -m robot.robot_control

Live model

  • To run a live model, ensure that checkpoints are saved in a format resembling checkpoints/<config>_<index>/model_<epoch>.pth. Then run:
  • Run the live model: python -m prediction.live_model --config <yaml file name from /config> --view --index <index> --epoch <epoch> (You do not need to specify index and epoch if using the provided models)

Demo

  • python -m demo.grasp_demo --config <yaml file name from /config> --view --index <index> --epoch <epoch>

Citing this work

If you use ViPER for academic research, please cite our paper. Below is the BibTeX entry for citation:

@inproceedings{collins2023visual,
  title={Visual Contact Pressure Estimation for Grippers in the Wild},
  author={Collins, Jeremy A and Houff, Cody and Grady, Patrick and Kemp, Charles C},
  booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year={2023}
  organization={IEEE}
}

About

We present Visual Pressure Estimation for Robots (ViPER), a neural network capable of visually estimating contact pressure for a soft gripper in the wild using only an RGB image from an eye-in-hand camera. ViPER enables precision grasping of small objects (e.g., a paperclip) on unseen cluttered surfaces.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages