Back to products Grasp / Manipulation Model

GraspGen

Name: GraspGen
Brand: NVIDIA

Community

NVIDIA's diffusion-based 6-DOF grasp generator: give it a segmented object point cloud or mesh, get back ranked gripper poses with confidence scores. Checkpoints for Franka Panda, Robotiq 2F-140 and a 30 mm suction cup. Research use only.

licenseResearch / Non-commercialframeworkPyTorch

$git clone https://github.com/NVlabs/GraspGen.git && cd GraspGen && pip install -e . && ./install_pointnet.sh

Input & output

Input

point cloud

A point cloud of one segmented object (N x 3), or an object mesh (.obj, .stl, .ply, .usd). Scene point clouds with instance segmentation supported, with optional point-cloud collision filtering.

Output

6-DOF grasp pose

Up to K grasps (top-k defaults to 100): gripper base-link poses as 4x4 transforms, plus a confidence score per grasp. Can also be written back into USD for Isaac Sim.

Example

Python

From grasp_gen/grasp_server.py. Input: a segmented object point cloud. Output: 4x4 gripper poses plus a confidence score each. Needs a CUDA GPU; the ZMQ server mode moves the GPU off the client.

from grasp_gen.grasp_server import GraspGenSampler, load_grasp_cfg

# One checkpoint per gripper: graspgen_robotiq_2f_140 / _franka_panda / _single_suction_cup_30mm
cfg = load_grasp_cfg("GraspGenModels/checkpoints/graspgen_robotiq_2f_140.yml")
sampler = GraspGenSampler(cfg)

# object_pc: (N, 3) point cloud of ONE segmented object (or points sampled from a mesh)
grasps, grasp_conf = GraspGenSampler.run_inference(
    object_pc, sampler, num_grasps=200, topk_num_grasps=100)

# grasps:     (K, 4, 4) gripper base-link poses
# grasp_conf: (K,) discriminator confidence per grasp

Running it

Runtime requirements: Linux + NVIDIA GPU with CUDA. pip/uv install pins Python 3.10 + PyTorch 2.1.0 (cu121) and builds a PointNet++ CUDA extension (needs CUDA headers + a C++ compiler). Docker images for CUDA 12.1 and 12.8. ZMQ server clients need no GPU. Retraining for a new gripper needs its URDF + a grasp dataset.
Model weights: huggingface.co/adithyamurali/GraspGenModels

Save it, compare it or record that you have used it — with an account.

Updated6d ago

At a glance

Approach

learned

Input Modality

point cloud

Output

6-DOF grasp pose

Overview

GraspGen (ICRA 2026) generates 6-DOF grasps for a single object with a Diffusion Transformer, then scores and filters them with a discriminator trained using an on-generator recipe. It takes a segmented object point cloud or an object mesh (.obj/.stl/.ply/.usd) and returns gripper base-link poses as 4x4 transforms, each with a confidence score; scene point clouds are supported with optional collision filtering. Released checkpoints cover three grippers (Franka Panda, Robotiq 2F-140, single 30 mm suction cup), trained on 57 million+ simulated grasps over 8,515 Objaverse objects. It can run in-process (Python), as a ZMQ inference server whose clients need no GPU, or as an MCP tool for LLM agents. On a real robot it needs an instance-segmentation front end (e.g. SAM 2) and a motion planner (e.g. cuRobo).

Key features

Point cloud or mesh in, ranked 6-DOF grasp poses + scores out
Checkpoints for Franka Panda, Robotiq 2F-140 and a 30 mm suction cup
State-of-the-art on the FetchBench grasping benchmark (ICRA 2026 paper)
ZMQ server and MCP tool modes — clients need no CUDA
57M+ grasp dataset released on Hugging Face (CC BY 4.0)

What it needs

Derived from its documented connections — mechanical, electrical, software.

InterfaceObject mesh · alternative to the segmented point cloud
Who provides it
InterfacePoint cloud · scope object · one segmented object; an object mesh works instead
Who provides it
RuntimeCUDA GPU runtime · version >=12.1 · pip/uv path pins torch 2.1.0+cu121 and compiles a PointNet++ CUDA extension; Docker images for CUDA 12.1 and 12.8
Who provides it
RuntimeLinux runtime · Docker images built on NVIDIA NGC PyTorch containers (23.07 / 25.03)
Who provides it

Alternatives

3 other Grasp / Manipulation Model products, closest hero specs and price first. Suggestions, not recommendations.

Compare in a table

All of them:Approach learned

ProductInput ModalityOutput

GraspGen

This product

Input Modalitypoint cloud

Output6-DOF grasp pose

This product is part of the Birdwave Atlas community.

Contribute evaluations, share integrations and help improve compatibility data.

Community feed

Discussions, evaluations, integrations and updates

Specifications

General

Approach

learned

Input Modality

point cloud

Output

6-DOF grasp pose

Framework

PyTorch

License

Research / Non-commercial

Back to products Grasp / Manipulation Model

GraspGen

NVIDIA

Community

licenseResearch / Non-commercialframeworkPyTorch

$git clone https://github.com/NVlabs/GraspGen.git && cd GraspGen && pip install -e . && ./install_pointnet.sh

Input & output

Input

point cloud

A point cloud of one segmented object (N x 3), or an object mesh (.obj, .stl, .ply, .usd). Scene point clouds with instance segmentation supported, with optional point-cloud collision filtering.

Output

6-DOF grasp pose

Up to K grasps (top-k defaults to 100): gripper base-link poses as 4x4 transforms, plus a confidence score per grasp. Can also be written back into USD for Isaac Sim.

Example

Python

From grasp_gen/grasp_server.py. Input: a segmented object point cloud. Output: 4x4 gripper poses plus a confidence score each. Needs a CUDA GPU; the ZMQ server mode moves the GPU off the client.

from grasp_gen.grasp_server import GraspGenSampler, load_grasp_cfg

# One checkpoint per gripper: graspgen_robotiq_2f_140 / _franka_panda / _single_suction_cup_30mm
cfg = load_grasp_cfg("GraspGenModels/checkpoints/graspgen_robotiq_2f_140.yml")
sampler = GraspGenSampler(cfg)

# object_pc: (N, 3) point cloud of ONE segmented object (or points sampled from a mesh)
grasps, grasp_conf = GraspGenSampler.run_inference(
    object_pc, sampler, num_grasps=200, topk_num_grasps=100)

# grasps:     (K, 4, 4) gripper base-link poses
# grasp_conf: (K,) discriminator confidence per grasp

Running it

Runtime requirements: Linux + NVIDIA GPU with CUDA. pip/uv install pins Python 3.10 + PyTorch 2.1.0 (cu121) and builds a PointNet++ CUDA extension (needs CUDA headers + a C++ compiler). Docker images for CUDA 12.1 and 12.8. ZMQ server clients need no GPU. Retraining for a new gripper needs its URDF + a grasp dataset.
Model weights: huggingface.co/adithyamurali/GraspGenModels

Save it, compare it or record that you have used it — with an account.

Updated6d ago

At a glance

Approach

learned

Input Modality

point cloud

Output

6-DOF grasp pose

Overview

Key features

Point cloud or mesh in, ranked 6-DOF grasp poses + scores out
Checkpoints for Franka Panda, Robotiq 2F-140 and a 30 mm suction cup
State-of-the-art on the FetchBench grasping benchmark (ICRA 2026 paper)
ZMQ server and MCP tool modes — clients need no CUDA
57M+ grasp dataset released on Hugging Face (CC BY 4.0)

What it needs

Derived from its documented connections — mechanical, electrical, software.

InterfaceObject mesh · alternative to the segmented point cloud
Who provides it
InterfacePoint cloud · scope object · one segmented object; an object mesh works instead
Who provides it
RuntimeCUDA GPU runtime · version >=12.1 · pip/uv path pins torch 2.1.0+cu121 and compiles a PointNet++ CUDA extension; Docker images for CUDA 12.1 and 12.8
Who provides it
RuntimeLinux runtime · Docker images built on NVIDIA NGC PyTorch containers (23.07 / 25.03)
Who provides it

Alternatives

3 other Grasp / Manipulation Model products, closest hero specs and price first. Suggestions, not recommendations.

Compare in a table

All of them:Approach learned

ProductInput ModalityOutput

GraspGen

This product

Input Modalitypoint cloud

Output6-DOF grasp pose

This product is part of the Birdwave Atlas community.

Contribute evaluations, share integrations and help improve compatibility data.

Community feed

Discussions, evaluations, integrations and updates

Discussions and reviews are for members

Specifications

General

Approach

learned

Input Modality

point cloud

Output

6-DOF grasp pose

Framework

PyTorch

License

Research / Non-commercial