Open Source RoboticsRobots, software & learning

AnyGrasp SDK

Robotics model

7-DoF parallel-jaw grasp detection and temporal grasp tracking SDK from point clouds, distributed as compiled Python binary libraries with machine-bound request-only license keys and checkpoints. [1] [2] [4]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Robotic grasping researchers and integrators willing to register machine-bound SDK licenses via upstream application form [1] [2]
Parameters
Not disclosed (compiled closed-source gsnet.so / tracker.so binaries and request-only checkpoints) [1] [2] [6]
Embodiments
Parallel-jaw grippers (max_gripper_width in (0.0, 0.1] m [4] [5]
Source rights
Some licenses are unresolved. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
License
not fully assessed for software, datasets, documentation · Details by part

Requirements · Research details and dates

On this page

Overview

AnyGrasp SDK (graspnet/anygrasp_sdk, IEEE T-RO 2023, arXiv:2212.08333) is a 7-DoF parallel-jaw grasp detection and temporal tracking SDK from Shanghai Jiao Tong University's MVIG Lab. Unlike open-source and open-weights grasp models, AnyGrasp is distributed as compiled closed-source Python C-extension binaries (gsnet.so and tracker.so) paired with machine-bound cryptographic license keys and request-only checkpoints. At pinned commit b8eaafc9eca7babd5208e7a5ade3c561060be4c5, the repository remains actively updated—adding steered grasping, Python 3.14 and CUDA 13 support, a redesigned license tool, and experimental aarch64 testing on the dev branch. [1] [2] [3] [4]

Model card summary

Architecture

Sparse 3D convolutional grasp perception and temporal tracking SDK built on MinkowskiEngine and PointNet++ (packaged as compiled Python C-extensions gsnet.so and tracker.so). Incorporates center-of-mass awareness and spatial-temporal grasp correspondence to output dense 7-DoF parallel-jaw grasps.

[1] [4] [5] [6]

Parameter count

Not disclosed (compiled closed-source gsnet.so / tracker.so binaries and request-only checkpoints)

Core network source code and checkpoint files are not publicly downloadable without license registration.

[1] [2] [6] · Editorial assessment

Input modalities
point-cloudrgbd

Unorganized 3D point clouds (derived from RGB-D cameras) plus optional steering regions and approach vectors.

[1] [5]

Output modalities
6d-posegripper-command

Dense 7-DoF parallel-jaw grasp poses (6-DoF pose plus gripper opening width up to 0.10 m) and temporal grasp correspondences.

[4] [5]

Action space

Grasp-pose perception and tracking SDK (no arm trajectory controller); predicts 7-DoF parallel-jaw grasp poses (6-DoF pose + gripper width <= 0.10 m) via create_detector(config) and detector.get_grasp(points).

[4] [5]

Trained / supported embodiments
  • Parallel-jaw grippers (max_gripper_width in (0.0, 0.1] m; embodiment-agnostic robot arm via external motion planner)

[4] [5]

Training datasets
  • Dense real-perception and analytic spatial-temporal grasp labels built upon GraspNet-1Billion / GraspNet datasets (per arXiv:2212.08333 and README dataset link)

[1] [4]

Interfaces
  • Python SDK API (gsnet.create_detector, detector.get_grasp, tracker.so)
  • Machine-ID license validation API (gsnet.get_feature_id, gsnet.check_license)

[2] [5] [6]

Checkpoints

CheckpointFormatSizeAccessPinned revision
AnyGrasp Detection checkpoint (request-only via license registration, placed in grasp_detection/log/)pytorchNot statedrequest-onlymain / b8eaafc9eca7
AnyGrasp Tracking checkpoint (request-only via license registration, placed in grasp_tracking/log/)pytorchNot statedrequest-onlymain / b8eaafc9eca7

Checkpoints and machine-bound license bundles are distributed only after manual approval of the Google Form application (HTTP 200 verified); no public direct-download URL exists.

[1] [2] [6]

Is this for you?

Use cases
  • Dense 7-DoF parallel-jaw grasp detection from unorganized 3D point clouds via gsnet.so (including region- and approach-vector-steered grasping)
  • Dynamic temporal grasp tracking on moving objects via tracker.so
  • High-throughput cluttered bin-picking pipelines paired with external robot motion planning

[1] [4] [5] [6]

Known limitations
  • Not open-source or open-weights: due to IP restrictions stated in README, core detector (gsnet.so) and tracker (tracker.so) are compiled closed-source binaries, and there is no root LICENSE file.
  • Model checkpoints and machine-bound license files (.lic, .public_key, .signature, licenseCfg.json) require submitting a machine feature ID via Google Form with ~5 workday turnaround.
  • Training code is not included; upstream points users wanting inspectable source code to graspnet/graspnet-baseline or graspnet/graspness_unofficial.
  • Requires compiling a custom fork of MinkowskiEngine (v0.5.4) and pointnet2 CUDA extensions.

[1] [2] [3] [6]

Safety and limitations
  • No open-source license file exists in the repository; execution is gated by a proprietary machine-bound cryptographic license key obtained via Google Form approval.
  • Core inference networks inside gsnet.so and tracker.so are opaque compiled binaries that cannot be audited or modified directly.
  • Outputs parallel-jaw grasp poses only—no robot kinematics, trajectory execution, or safety controller is included.
  • Open Source Robotics has not applied for a license key, compiled MinkowskiEngine, downloaded checkpoints, run inference, or tested AnyGrasp on physical hardware.

[1] [2] [3] [6]

Compute & fine-tuning

Inference compute
  • Documented hardware: NVIDIA CUDA GPU (CUDA 11.x, 12.x, or 13.x required for PyTorch 1.7.1+, MinkowskiEngine v0.5.4, and pointnet2; specific VRAM minimum not stated)

README lists CUDA 11.x/12.x/13.x and dev-branch aarch64 SDK testing, without stating a scalar VRAM minimum or latency.

[1]

Fine-tuning
  • Supported: No

Inference-only compiled SDK binaries; training/fine-tuning code is not released in anygrasp_sdk.

[1] [6]

Host computer

Linux x86_64 host with NVIDIA CUDA GPU (CUDA 11.x/12.x/13.x); experimental aarch64 SDK is under testing on the dev branch.

[1] [2]

Operating system
  • Linux

Compiled binaries target x86_64-linux-gnu (plus experimental aarch64 Linux on dev branch).

[1] [2] [6]

Dependencies
  • Python 3.6-3.14 and PyTorch 1.7.1+ with CUDA 11.x/12.x/13.x
  • Modified MinkowskiEngine v0.5.4 (chenxi-wang/MinkowskiEngine) with OpenBLAS
  • Compiled pointnet2 module and graspnetAPI
  • Compiled gsnet.so / tracker.so binaries and machine-bound license folder

[1] [2] [6]

Skills
  • Linux C++/CUDA compilation (MinkowskiEngine and PointNet++)
  • 3D point-cloud pipelines and external robot arm motion planning

Inferred from MinkowskiEngine/pointnet2 build steps and Python SDK usage.

[1] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [1] · Model card / documentation [1]

  • Install PyTorch 1.7.1+ with CUDA 11.x/12.x/13.x and compile the modified MinkowskiEngine v0.5.4 fork (chenxi-wang/MinkowskiEngine) with OpenBLAS.
  • Install pip requirements (requirements.txt), compile the bundled pointnet2 module, and install graspnet/graspnetAPI.
  • Copy the matching Python-version gsnet.*.so / tracker.*.so binary, run get_feature_id(), and submit the machine feature ID via the Google Form.
  • Unzip the approved license folder and place the provided checkpoints under grasp_detection/log/ and grasp_tracking/log/, then validate via check_license('license') and run demo.sh.

[1] [2] [6]

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
Real-World Cluttered Bin Clearing (>300 unseen objects; IEEE T-RO 2023, arXiv:2212.08333)93.3% grasp success rate and >900 mean-picks-per-hour on a single-arm system (reported on par with human subjects under controlled conditions); also demonstrates dynamic catching of swimming robot fish.Upstream-reported hardware evaluation in arXiv:2212.08333 abstract; OSR has not obtained a license key or reproduced any experiments.

[4]

Project notes

Closed binary SDK, API, and dependencies

The public repository provides demo wrappers, pointnet2 build code, and version-specific compiled gsnet.*.so and tracker.*.so binaries for Linux x86_64. In grasp_detection/USAGE.md, inference is invoked via create_detector(config) (taking checkpoint_path, max_gripper_width in (0.0, 0.1] meters, and gripper_height) and detector.get_grasp(points) over unorganized 3D point clouds, with July 2026 updates adding region and approach-vector steering. No training or fine-tuning code is released; upstream explicitly directs users who want inspectable network source code to graspnet/graspnet-baseline or graspnet/graspness_unofficial. [1] [5] [6]

Installation requires PyTorch 1.7.1+ with CUDA 11.x, 12.x, or 13.x, compiling a modified fork of MinkowskiEngine v0.5.4 (chenxi-wang/MinkowskiEngine) with OpenBLAS, building the bundled pointnet2 module, and installing graspnetAPI. Exact parameter counts, checkpoint file sizes, and minimum GPU VRAM are not established by the supplied documentation. [1] [6]

Upstream-reported grasping performance

In arXiv:2212.08333, the authors report a 93.3% grasp success rate clearing cluttered bins containing over 300 unseen objects and throughput exceeding 900 mean-picks-per-hour on a single-arm robot system, alongside dynamic tracking demonstrations catching swimming robot fish. All results are upstream-reported; Open Source Robotics has not applied for a license key, compiled MinkowskiEngine, downloaded checkpoints, run training, fine-tuning, or inference, or tested AnyGrasp on physical hardware. [4]

What we checked

Source researched only; no license registration, installation, inference, benchmark, or physical hardware test performed. No affiliate links or reused media.

Not yet established

  • Documented closed-loop control frequency (Hz)

Updates

Upstream last updated
Jul 13, 2026 [3]
Latest release
Not established
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
commit snapshot; no GitHub release / b8eaafc9eca7babd5208e7a5ade3c561060be4c5 · software repository

Separate license matrix

Include AnyGrasp SDK with explicit partial/closed-binary warnings because it is widely referenced alongside open grasping models, while making clear that core binaries are closed-source, there is no root open-source license, and weights/keys are request-only.

LayerAvailability / license
Codeavailable · License not fully assessed
Demo/wrapper Python scripts and pointnet2 source are public on GitHub with no root LICENSE file; core gsnet.so and tracker.so inference engines are closed-source compiled binaries requiring a machine-bound license key. [1] [2] [3]
Model weightsunavailable · License not fully assessed
Weights are request-only and distributed privately alongside machine-bound license files after submitting a Google Form application (HTTP 200 verified); no public weights download or open license exists. [1] [2] [6]
Training datasetsavailable · License not fully assessed
README links to graspnet.net/datasets.html; training code and AnyGrasp's full proprietary training pipeline are not included in this repository. [1]
Documentationavailable · License not fully assessed
Repository README, USAGE.md, CHANGELOG.md, and license_registration/README.md; no root LICENSE file. [1] [3] [5]

Request-only license keys and lack of open-source license

The repository contains no root LICENSE file. Upstream states that "due to the IP issue," the SDK library can only be released in a licensed manner: users must extract a machine feature ID (gsnet.get_feature_id()), submit it via a Google Form (verified HTTP 200; ~5 workday reply window), and place the returned license/ folder (licenseCfg.json, .public_key, .signature, and .lic) directly inside each demo directory (grasp_detection/ and grasp_tracking/), with model checkpoints placed inside each directory's log/ subfolder before running demo.sh. Approval criteria, commercial-use eligibility, and any applicable fees are not established in the supplied repository materials, so integrators should confirm terms directly with upstream before committing to deployment. [1] [2] [3] [6]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history

2026-11-21 — Initial 0.8.0-draft.1 model profile for AnyGrasp SDK (partial/closed binary SDK, request-only license keys and checkpoints).