Open Source RoboticsRobots, software & learning

UniDexGrasp

Robotics model

Two-stage universal dexterous grasping framework for ShadowHand combining probabilistic grasp proposal generation (GraspIPDF, GraspGlow, ContactNet) with an Isaac Gym goal-conditioned RL and DAgger vision policy. [1] [2] [3] [5]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Dexterous manipulation and RL researchers reproducing or extending CVPR 2023 ShadowHand grasping benchmarks in Isaac Gym [1] [3]
Parameters
Not stated in upstream READMEs or arXiv:2303.00938 abstract (multi-module pipeline: GraspIPDF, GraspGlow,… [2] [3] [5]
Embodiments
ShadowHand (simulated dexterous robotic hand in Isaac Gym / MJCF) [2] [3]
Source rights
Named licenses are documented by part. Includes noncommercial terms. Review licenses

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

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

UniDexGrasp (PKU-EPIC/UniDexGrasp, CVPR 2023, arXiv:2303.00938) is a two-stage universal tabletop dexterous grasping framework for the ShadowHand from Peking University's EPIC Lab. At pinned commit 36c9bfcf7cedc987dd15e2384c548e51d5aadc21, the repository provides source code for both grasp proposal generation (dexgrasp_generation) and goal-conditioned RL/DAgger policy learning (dexgrasp_policy), alongside a PKU dataset mirror—though no pretrained model checkpoint links were found in the checked repository READMEs or project page as of 2026-11-21. [1] [2] [3] [4] [6]

Model card summary

Architecture

Two-stage dexterous grasping architecture: Stage 1 (dexgrasp_generation) factorizes grasp pose generation into GraspIPDF (Implicit PDF over SO(3) rotation), GraspGlow (normalizing flow via nflows over translation and hand articulation), and ContactNet (PointNet++ contact refinement); Stage 2 (dexgrasp_policy) trains a goal-conditioned PPO teacher policy with state canonicalization and object curriculum in Isaac Gym, then distills it via DAgger into a PointNet++ vision-based execution policy.

[2] [3] [5]

Parameter count

Not stated in upstream READMEs or arXiv:2303.00938 abstract (multi-module pipeline: GraspIPDF, GraspGlow, ContactNet, and PointNet++/MLP PPO+DAgger policies; no pretrained checkpoints released)

Parameter counts are omitted from upstream READMEs and abstract, and no pretrained checkpoints are distributed.

[2] [3] [5] · Editorial assessment

Input modalities
point-cloudproprioception

Stage-1 and Stage-2 vision policy take tabletop object point clouds (plus hand proprioception and target grasp goal pose); Stage-2 teacher uses simulator oracle states.

[3] [5]

Output modalities
6d-posejoint-positions

Stage-1 outputs 6-DoF wrist pose and ShadowHand articulation proposals; Stage-2 outputs closed-loop ShadowHand joint/wrist execution actions.

[2] [3] [5]

Action space

Two-stage dexterous control: Stage 1 generates diverse 6-DoF wrist poses and ShadowHand joint articulation proposals; Stage 2 executes closed-loop goal-conditioned ShadowHand actions in Isaac Gym.

[2] [3] [5]

Trained / supported embodiments
  • ShadowHand (simulated dexterous robotic hand in Isaac Gym / MJCF)

[2] [3]

Training datasets
  • UniDexGrasp synthesized dexterous grasp dataset (DFCData + ShadowHand MJCF on PKU mirror; 3,200 training object instances, 141 seen-category test instances, 100 unseen-category test instances)

[2] [3] [5]

Interfaces
  • Python training/eval CLIs (dexgrasp_generation/network/train.py, eval.py)
  • Bash RL/DAgger launchers (run_train_ppo_state.sh, run_train_ppo_vision.sh, run_train_dagger.sh)

[2] [3]

Checkpoints

CheckpointFormatSizeAccessPinned revision
UniDexGrasp Stage-1 Grasp Proposal Networks (GraspIPDF, GraspGlow, ContactNet — train from scratch)pytorchNot statedweights-not-releasedNot pinned
UniDexGrasp Stage-2 Goal-Conditioned Execution Policy (PPO teacher + DAgger vision student — train from scratch)pytorchNot statedweights-not-releasedNot pinned

Neither the GitHub READMEs nor the project page link pretrained weight checkpoints; the PKU mirror provides dataset/MJCF assets for training from scratch.

[2] [3] [6]

Is this for you?

Use cases
  • Training Stage-1 probabilistic dexterous grasp proposal generators (GraspIPDF, GraspGlow, ContactNet) on DFCData
  • Training Stage-2 ShadowHand state-based PPO teachers and distilling point-cloud vision policies via DAgger in Isaac Gym
  • Benchmarking universal tabletop dexterous grasping across 3,200 training objects, 141 seen-category test objects, and 100 unseen-category test objects

[2] [3] [5]

Known limitations
  • No pretrained model checkpoints are released in the GitHub repository or on the PKU dataset mirror; users must train both Stage-1 proposal networks and Stage-2 RL/DAgger policies from scratch.
  • Licensed under CC-BY-NC-4.0 (declared in root README with no standalone root LICENSE file), restricting use to non-commercial purposes.
  • Depends on legacy software stacks: Ubuntu 20.04, Python 3.8, PyTorch 1.10 / CUDA 11.3, PyTorch3D 0.6.2, and NVIDIA Isaac Gym Preview Release 3.
  • Simulation-only tabletop ShadowHand pipeline; no real-robot hardware deployment code is included.

[1] [2] [3] [4] [6]

Safety and limitations
  • CC-BY-NC-4.0 notice in README prohibits commercial use; there is no separate root LICENSE file in the repository.
  • Pretrained weights are not released, requiring full simulation training from scratch before evaluation.
  • Evaluated and implemented only in Isaac Gym Preview Release 3 for a simulated ShadowHand; no physical robot safety controllers or sim-to-real hardware wrappers are provided.
  • Open Source Robotics has not downloaded the PKU dataset mirror, installed Isaac Gym, trained models, or run inference.

[1] [2] [3] [4] [6]

Compute & fine-tuning

Inference compute
  • Documented hardware: NVIDIA CUDA GPU required (CUDA 11.3 + Isaac Gym Preview Release 3 + PyTorch3D 0.6.2 + pointnet2_ops; exact VRAM minimum not stated)
  • Edge / consumer capable: No

CUDA 11.3 and NVIDIA Isaac Gym GPU requirements stated in installation instructions; scalar VRAM minimum and latency are omitted.

[2] [3] · Editorial assessment

Fine-tuning
  • Supported: Yes
  • Methods: Stage-1 training from scratch (GraspIPDF, GraspGlow, ContactNet via dexgrasp_generation/network/train.py); Stage-2 PPO state/vision RL training and DAgger state-to-vision distillation via dexgrasp_policy/dexgrasp/script/run_train_*.sh
  • Recipe: Upstream fine-tuning documentation

[2] [3]

Host computer

Ubuntu 20.04 host with NVIDIA CUDA 11.3-compatible GPU capable of running Isaac Gym Preview Release 3; exact VRAM requirements are not stated in READMEs.

[2] [3]

Operating system
  • Linux

dexgrasp_generation/README.md explicitly states tested on Ubuntu 20.04 and installs linux-64 PyTorch3D; Isaac Gym targets Linux.

[2] [3]

Dependencies
  • Ubuntu 20.04, Python 3.8, PyTorch 1.10.0, CUDA 11.3
  • PyTorch3D 0.6.2, pytorch_kinematics, nflows, CSDF (Stage 1)
  • NVIDIA Isaac Gym Preview Release 3 and Pointnet2_PyTorch pointnet2_ops_lib (Stage 2)

[2] [3]

Skills
  • Isaac Gym Preview Release 3 and CUDA 11.3 / PyTorch 1.10 environment setup
  • Reinforcement learning (PPO), DAgger imitation distillation, and 3D point-cloud networks

Inferred from Stage-1 and Stage-2 installation and training READMEs.

[2] [3] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [2] · Model card / documentation [1]

  • For Stage 1 (dexgrasp_generation), create a Python 3.8 conda env on Ubuntu 20.04 with PyTorch 1.10.0 / CUDA 11.3, PyTorch3D 0.6.2, pytorch_kinematics, nflows, and CSDF.
  • Download the ShadowHand MJCF and DFCData grasp labels from the PKU mirror into dexgrasp_generation/data/ and train GraspIPDF, GraspGlow, and ContactNet.
  • For Stage 2 (dexgrasp_policy), create a Python 3.8 conda env, install NVIDIA Isaac Gym Preview Release 3, install dexgrasp and pointnet2_ops_lib.
  • Run script/run_train_ppo_state.sh, run_train_ppo_vision.sh, or run_train_dagger.sh in dexgrasp_policy/dexgrasp/.

[2] [3]

Upstream-reported benchmarks

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

BenchmarkUpstream-reported scoreScope & caveats
UniDexGrasp Tabletop Dexterous Grasping in Isaac Gym (CVPR 2023, arXiv:2303.00938)Over 60% average grasp success rate across thousands of tabletop object instances (seen and unseen categories) with minimal generalization gap.Upstream-reported simulation evaluation in arXiv:2303.00938 abstract using 3,200 train / 141 seen-category test / 100 unseen-category test object splits; OSR has not reproduced training or evaluation.

[3] [5]

Project notes

Two-stage architecture and legacy simulation stack

UniDexGrasp decouples dexterous grasping from raw object point clouds into two stages. Stage 1 (dexgrasp_generation) factorizes grasp pose synthesis into three modules trained on the synthesized DFCData dataset and ShadowHand MJCF assets: GraspIPDF (Implicit PDF over SO(3) rotations), GraspGlow (normalizing flow via nflows over wrist translation and finger articulation), and ContactNet (PointNet++ contact refinement). Stage 2 (dexgrasp_policy) replaces classical motion planning with a goal-conditioned RL execution policy in NVIDIA Isaac Gym, training a state-based PPO teacher with state canonicalization and an object curriculum before distilling it via DAgger into a PointNet++ vision-based student policy. [2] [3] [5]

Because no pretrained checkpoint links were found in the checked READMEs or project page as of 2026-11-21 (and mirror payloads were not downloaded or inspected), reproducing the full pipeline from those materials requires training both stages. Both sub-projects rely on legacy 2021–2023 environments: Stage 1 was tested on Ubuntu 20.04 with Python 3.8, PyTorch 1.10.0, CUDA 11.3, PyTorch3D 0.6.2, pytorch_kinematics, nflows, and CSDF, while Stage 2 requires Python 3.8, NVIDIA Isaac Gym Preview Release 3, and Pointnet2_PyTorch (pointnet2_ops_lib). The supplied READMEs do not specify minimum GPU VRAM or training wall-clock time, so the checked sources do not support a training compute or cost estimate. [2] [3] [6]

Upstream-reported simulation performance

In arXiv:2303.00938, the authors evaluate the integrated two-stage pipeline across tabletop splits comprising 3,200 training object instances, 141 seen-category test instances, and 100 unseen-category test instances, reporting an average grasp success rate above 60% with minimal generalization gap. All benchmark figures are upstream-reported in simulation; Open Source Robotics has not downloaded the PKU dataset mirror, installed Isaac Gym, run training, fine-tuning, or inference, or tested UniDexGrasp on hardware. [3] [5]

What we checked

Source researched only; no installation, training, fine-tuning, 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
May 25, 2024 [4]
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 / 36c9bfcf7cedc987dd15e2384c548e51d5aadc21 · software repository

Separate license matrix

Include UniDexGrasp as a foundational inspectable dexterous grasping training codebase and dataset release, while clearly flagging that pretrained weights are not released, licensing is CC-BY-NC-4.0, and the stack targets legacy Isaac Gym Preview Release 3.

LayerAvailability / license
Codeavailable · CC-BY-NC-4.0 (declared in root README; no root LICENSE file)
Root README states 'This work and the dataset are licensed under CC BY-NC 4.0'; no standalone root LICENSE file is present in the repository. [1] [4]
Model weightsnot found · License not fully assessed
No pretrained model checkpoints are published in the repository READMEs or project page; only training code and dataset/MJCF files are distributed. [2] [3] [6]
Training datasetsavailable · CC-BY-NC-4.0 (declared in root README)
PKU mirror URL verified HTTP 200 (provides ShadowHand MJCF and DFCData grasp labels); archive payloads were not downloaded. [1] [2] [6]
Documentationavailable · CC-BY-NC-4.0 (declared in root README)
Root README plus sub-READMEs in dexgrasp_generation/ and dexgrasp_policy/. [1] [2] [3]

Non-commercial license notice

The repository has no standalone root LICENSE file, but the root README.md explicitly states that "This work and the dataset are licensed under CC BY-NC 4.0," restricting code and data use to non-commercial purposes. The repository has had no commits since 2024-05-25 and no versioned GitHub releases. [1] [4]

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 UniDexGrasp (training code + dataset only, weights-not-released, CC-BY-NC-4.0).