Execution and Control

AnyGrasp

AnyGrasp is a robust grasp perception tool for predicting and tracking 6-DoF robotic grasps from RGB-D observations and point clouds.

Tool Introduction

Core parameters, trigger timing, and visual before/after demo references.

Short Explanation

Provide an RGB-D observation or point cloud, and AnyGrasp predicts feasible 6-DoF grasp poses with scores and gripper widths for robotic manipulation.

InputRGB-D image / point cloud

Output6-DoF grasp poses, scores, widths

Trigger TimingTriggered on demand from the source demo or local example command.

RuntimePython / PyTorch / MinkowskiEngine / SDK demo

BeforeRGB-D image / point cloud

Prepare the scene, image, video, sensor stream, prompt, or configuration expected by the original project.

After6-DoF grasp poses, scores, widths

Read the produced visualization, prediction, map, trajectory, mask, grasp pose, or other documented artifact.

Preset Example

A quick-run style example for the documentation page.

Inputtools/anygrasp/grasp_detection/example_data/color.png + depth.png

PromptFind stable grasps within the gripper width limit

ExpectedRanked 6-DoF grasp poses with visualization, score, width, and optional temporal tracking output.

Parameters And Output

Readable controls and the meaning of each returned artifact.

Parameter Explanation

color_imagefile

RGB frame aligned with the depth input.

depth_imagefile

Metric depth image used to recover 3D geometry for grasp candidates.

max_gripper_widthslider0.10 m

Filters grasps that exceed the physical gripper opening.

filterselectoneeuro

Temporal smoothing option for tracking grasps across frames.

Output Explanation

pose

6-DoF grasp frame describing gripper position and orientation.

score

Predicted grasp quality used to rank candidates.

width

Required gripper opening width for the selected grasp.

How To Use

Official resources, deployment steps, academic context, citation, and source-reported benchmark numbers.

Resources

GitHubhttps://github.com/graspnet/anygrasp_sdk Code Downloadhttps://github.com/graspnet/anygrasp_sdk/archive/refs/heads/main.zip Project Pagehttps://graspnet.net/anygrasp.html arXivhttps://arxiv.org/abs/2212.08333 GraspNet Datasethttps://graspnet.net/datasets.html

Deployment Notes

Register for the official AnyGrasp SDK and download the licensed binaries and checkpoints.
Install PyTorch, MinkowskiEngine, Open3D, and the repository dependencies expected by the SDK.
Prepare aligned RGB-D images or point clouds under the example data folder.
Run detection or tracking demos and store visualized grasps under tools/anygrasp/runs/.

Relative Path Example

# Relative-path local entry for the AnyGrasp tool folder
python tools/anygrasp/grasp_detection/demo.py   --checkpoint_path tools/anygrasp/log/checkpoint_detection.tar   --max_gripper_width 0.1   --gripper_height 0.03   --debug

# Optional temporal tracking entry:
python tools/anygrasp/grasp_tracking/demo.py   --checkpoint_path tools/anygrasp/log/checkpoint_tracking.tar   -filter oneeuro   --debug

# Suggested repository layout when adding local files:
# tools/anygrasp/README.md
# tools/anygrasp/grasp_detection/demo.py
# tools/anygrasp/grasp_tracking/demo.py
# tools/anygrasp/grasp_detection/example_data/color.png
# tools/anygrasp/grasp_detection/example_data/depth.png
# tools/anygrasp/log/

# This page documents the path. The static page does not execute AnyGrasp.
# The official implementation requires the licensed AnyGrasp SDK binaries and model weights.

Expected Result Shape

{
  "tool": "anygrasp",
  "status": "ok",
  "grasps": [
    {
      "label": "6-DoF grasp perception",
      "score": 0.87,
      "output": "6-DoF grasp poses, scores, widths"
    }
  ],
  "timing": {
    "runtime": "The paper reports 100 ms grasp prediction and less than 200 ms overall decision time; single UR5 setup reaches 900+ mean picks per hour.",
    "device": "documented in source benchmark when available"
  },
  "artifacts": {
    "visualization": "tools/anygrasp/runs/visualization.png",
    "raw_predictions": "tools/anygrasp/runs/predictions.json"
  }
}

Paper figure

Academic Info

Paper identity and contribution summary.

TitleAnyGrasp: Robust and Efficient Grasp Perception in Spatial and Temporal Domains

AuthorsHao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, Cewu Lu

VenueIEEE Transactions on Robotics, 2023 / arXiv:2212.08333

ContributionProvides a grasp perception pipeline that detects robust 6-DoF grasp candidates and supports temporal tracking, enabling robotic manipulation in cluttered and dynamic scenes.

Citation

@misc{anygrasp2023,
  title={AnyGrasp: Robust and Efficient Grasp Perception in Spatial and Temporal Domains},
  author={Hao-Shu Fang and Chenxi Wang and Hongjie Fang and Minghao Gou and Jirong Liu and Hengxu Yan and Wenhai Liu and Yichen Xie and Cewu Lu},
  year={2023},
  note={IEEE Transactions on Robotics, 2023 / arXiv:2212.08333},
  url={https://arxiv.org/abs/2212.08333}
}

Benchmark

Only compact, source-reported numbers are shown here.

Dataset	Metric	Value	Runtime	Source
Real bin-picking benchmark	Attempt-centric success	93.3% AnyGrasp vs 72.2% DexNet 4.0; object completion 99.8%	100 ms prediction, <200 ms decision time	T-RO 2023 paper
Dynamic fish catching	Success rate	75.5% AnyGrasp vs 62.5% heuristic baseline	Temporal grasp tracking	AnyGrasp paper

Artifacts

AnyGrasp paper, success-rate table, RGB-D examples, grasp detection demo, grasp tracking demo, SDK registration notes, and real-robot videos.

Demo Images

Visual references from the original tool. Click any image to inspect the original size.