Install¶

GraspPanda runs on Linux with Python 3.10 or newer and PyTorch 2.1 or newer. Installing takes a few minutes; checking it takes under a minute and needs no dataset.

1. Install PyTorch¶

Install PyTorch for your CUDA version, following pytorch.org.

2. Install GraspPanda¶

git clone https://github.com/Daeda1used/GraspPanda.git
cd GraspPanda
pip install -e .

There is no release on PyPI yet: pip install grasppanda arrives with the first one.

3. Add what your method needs¶

As shipped, GraspNet-baseline, Scale-Balanced-Grasp and HGGD need nothing more. GSNet, EconomicGrasp and FineGrasp use sparse convolution: add the spconv extra that matches the CUDA build of your PyTorch.

# prints your CUDA build: 11.8 means spconv-cu118
python -c "import torch; print(torch.version.cuda)"
pip install -e ".[spconv-cu118]"

A CPU-only PyTorch prints None: use spconv-cpu.

An option from another library, such as an image encoder from timm, needs its extra too. Its slot page names it, and a missing extra stops with the command for your install.

Every extra

Extra	Installs	Needed for
`spconv-cu118`, `spconv-cu121`, `spconv-cu124`, `spconv-cu126`, `spconv-cpu`	spconv for your CUDA build	GSNet, EconomicGrasp, FineGrasp, and the scene encoders built on sparse convolution (sparse-voxel networks, PTv3, LitePT)
`timm`, `transformers`, `torchvision`	a model library	image encoders, and the Uni3D scene encoder (`timm`)
`kornia`	Kornia	its losses
`torch_geometric`	PyG	the PyG scene encoder and grouper
`trimesh`, `roma`, `healpix`	a geometry library	gripper-mesh collision, some view stages
`ocnn`, `mamba-ssm`	octree and selective-scan operators	the OctFormer and PointMamba scene encoders
`optim`	timm, pytorch_optimizer, schedulefree	optimizers and schedules beyond `torch.optim`
`fpsample`, `torch-cluster`	sampling libraries	the seed samplers that bridge to them
`vis`	Open3D	the window of `grasppanda demo --show`; pictures need no extra
`tensorboard`, `wandb`	a logger	logging to that service
`dev`, `docs`	test, lint and docs tools	contributing, building this site

Extras combine: pip install -e ".[spconv-cu118,timm]". From an installed wheel the same extras install with pip install "grasppanda[spconv-cu118]".

spconv on a CPU-only machine

spconv-cpu (spconv 2.3.8) returns wrong, non-repeatable results when PyTorch runs more than one thread. GraspPanda therefore runs each sparse convolution of a CPU tensor on one thread and gives the other operators their threads back; a run's log and the probe report say so once. It also compares the first sparse convolution with a dense one and stops with an explanation if they differ. No setting is needed. A CUDA build of spconv computes CPU tensors correctly with any number of threads.

Compiled operators

Furthest point sampling, ball query, cylinder query and nearest neighbours have compiled kernels that are built on first use when a C++ compiler is present. Without one, the pure-PyTorch reference runs. See Operators and backends.

4. Check the install¶

grasppanda env
grasppanda train projects/example_project/configs/toy_train.yaml

env prints the interpreter, the PyTorch build, the GPUs and the optional backends; paste it into bug reports. train trains a toy detector on synthetic scenes, on the CPU, in under a minute. It ends with ✓ trained: everything is wired. The toy's grasps mean nothing.

What train prints (shortened)

ʕ•ᴥ•ʔ  GraspPanda <version>  ·  6-DoF grasp detection in clutter
  train
  config      projects/example_project/configs/toy_train.yaml  (hash be81c4b1301a)
  run         work_dirs/toy_train/20261018-120000
  device      cpu · 1 process · seed 0 · deterministic off
  model       ToyDetector · 7,601 parameters (7,601 trainable)
  data        32 frames · 8 batches of 4 per epoch
  schedule    5 epochs · Adam(lr=0.005, weight_decay=0.0) · ExponentialLR(gamma=0.95) per epoch
  evaluation  data.val every 1 epoch(s) · variants raw, centre · best by raw/AP
epoch 1/5  iter 8/8  loss 3.9372  loss/score 3.8366  loss/width 0.1006  ...
evaluation after epoch 1 (8 frames, 0.4 s)
...
✓ trained epochs 1-5 in 15.6 s
  checkpoint  work_dirs/toy_train/20261018-120000/checkpoints/epoch_0005.pth
  best        raw/AP = 91.3105 at epoch 5  (work_dirs/toy_train/20261018-120000/checkpoints/best.pth)

The run directory is named after the time it started. The losses and the AP of the toy differ from run to run on another machine.

Next¶

Run a method from its released weights, or swap a component.