GraspPanda

Open-source toolbox that runs six published 6-DoF grasp detectors in one pipeline, with swappable stages and one pinned evaluator per dataset.

GraspPanda: a toolbox for 6-DoF grasp detection in cluttered scenes

Strong grasp detectors each live in their own repository, with their own data pipeline and often their own copy of the evaluator, which makes them hard to compare. GraspPanda brings them into one pipeline with one pinned evaluator per dataset, and only advertises what has been verified.

  • Faithful ports. Six published methods in two families, each matching its reference code stage by stage: GraspNet-baseline (CVPR 2020), GSNet (ICCV 2021), Scale-Balanced-Grasp (CoRL 2022), HGGD (RA-L 2023), EconomicGrasp (ECCV 2024) and FineGrasp (2025).
  • Swap stages. Encoders, seed proposal, view, grouper, heads, losses, post-processing and the training recipe are slots, each replaced with one line: --set <slot>=<option>. Most incompatible combinations are refused before anything is built, with the reason.
  • Measured, not assumed. The documentation shows, for each option, under how many methods it runs and learns, and why the others are left out. GraspNet-1Billion and GraspClutter6D use pinned evaluators equal to the official tools.
GraspPanda pipeline: two detector families and their swappable slots
Five methods share the seed-based skeleton and HGGD forms its own family; post-processing and the training recipe are shared by all six.

GraspPanda is released under the Apache-2.0 licence and is in alpha.