Skip to content

Run a method

Every ported method runs from its released weights: download the file, convert it once, and get grasps on the example scene in minutes. No dataset is needed.

Which method first?

GraspNet-baseline, Scale-Balanced-Grasp and HGGD need no extra install. HGGD reads a colour image registered to the depth image as well; the others read depth alone.

Method Reads Extra Weights
from		 Licence Demo
GraspNet-baseline depth none Google Drive non-commercial 10.5 s
GSNet depth spconv Google Drive non-commercial 14.8 s
Scale-Balanced-Grasp depth none Google Drive MIT, with non-commercial parts 10.5 s
EconomicGrasp depth spconv GitHub release MIT, with non-commercial parts 14.0 s
FineGrasp depth spconv Hugging Face Apache-2.0 14.8 s
HGGD RGB-D none Tsinghua Cloud MIT 14.5 s
  • Extra: spconv is installed as the extra that matches your PyTorch: spconv-cu118, spconv-cu121, spconv-cu124, spconv-cu126, spconv-cpu.
  • Licence: hover for the full text; each method page states it in full.
  • Demo is the whole grasppanda demo … --example command, measured on one NVIDIA RTX A6000 shared with other jobs; no run was timed on a CPU.
  • Every checkpoint was trained on GraspNet-1Billion, whose data and labels are under CC BY-NC-SA 4.0.
The example scene

The example scene: the colour image and the depth image of six primitives on a table

--example is a synthetic table-top scene that ships with GraspPanda: six geometric primitives on a table, ray-cast into a 1280 × 720 depth image in millimetres, with a colour image, the camera matrix and an object mask.

  • It is computed from formulas, so it needs no download and carries no dataset licence.
  • It is not a camera image: surfaces are perfectly smooth, with no noise, no missing depth and no reflection. Every method was trained on real scenes, so the example is outside its training distribution.
  • It has no grasp labels, so nothing checks the quality of the predicted grasps. It shows that the installation and the weights work end to end.

Pick a method

Each page walks you through it: the weights and their licence, the conversion, the demo, the same from Python and in the browser, and a first number on the benchmark.

  • GraspNet-baseline · CVPR 2020 · seed-based

    The benchmark's own baseline: a PointNet++ encoder, approach views and cylinder crops.

    Run it →

  • GSNet · ICCV 2021 · seed-based

    Finds the points worth grasping first, on a sparse U-Net encoder.

    Run it →

  • Scale-Balanced-Grasp · CoRL 2022 · seed-based

    Balances grasps of every width, so thin objects are not drowned out.

    Run it →

  • EconomicGrasp · ECCV 2024 · seed-based

    Trains on a compact set of labels with an interaction grouper.

    Run it →

  • FineGrasp · arXiv 2025 · seed-based

    Builds on EconomicGrasp with crops at several scales, for delicate objects.

    Run it →

  • HGGD · RA-L 2023 · own family

    A heat-map on the colour and depth images guides where to look for grasps.

    Run it →

On your camera

A depth image and its intrinsics are enough. For the seed-based methods, --workspace crops the cloud to the objects, as they were trained; HGGD reads the whole image and crops no workspace.

grasppanda demo <method config> --weights <converted file> \
    --depth depth.png --intrinsics fx,fy,cx,cy \
    --workspace -0.25 -0.2 0.35 0.25 0.2 0.8

HGGD also reads the colour image registered to the depth image: add --rgb rgb.png and leave out --workspace. The methods were trained on GraspNet-1Billion (RealSense D435 and Azure Kinect); results on other cameras have not been measured. Units, the depth scale and the usual mistakes: Your own camera.

From Python or in the browser

  • From Python: GraspInferencer runs the demo's pipeline and returns the grasps.
  • In the browser: grasppanda explore --weights-dir <folder> runs a method on the example scene or an uploaded depth image.

Next

Swap a component of the method you ran.