VNB-Grasp: Robust Dexterous Grasping under Multimodal Uncertainty

VNB-Grasp (Overview and VNB-Informed MPC Rollouts). Method overview, simulation evaluation, and hardware grasp execution.

A differentiable Gaussian-mixture belief represents uncertainty in object pose and latent contact parameters, enabling pathwise gradients of a smooth Conditional Value-at-Risk (CVaR) surrogate for risk-sensitive grasp planning.

Reusable Components

Differentiable Belief Representation

Gaussian mixtures retain multiple hypotheses about contact conditions and object pose. Gumbel–Softmax component selection and location–scale reparameterization make samples differentiable with respect to the belief parameters, allowing a smooth CVaR surrogate to guide optimization of adverse grasp outcomes.

Explore the Belief Module

MuJoCo Grasp Robustness Benchmark

The benchmark evaluates friction sensitivity and perturbation robustness across nominal, adversarial, wide, and bimodal friction regimes. Post-grasp tests apply lateral impulses, torque impulses, and sudden friction drops to measure grasp success and perturbation survival.

Explore the Benchmark Configuration

Simulation and Hardware Results

The paper’s Table III reports means over objects, risk levels, and seeds in three friction regimes: nominal, wide, and bimodal. The released benchmark additionally includes the adversarial regime. Robust success requires nominal success and survival of at least 50% of the perturbation tests.

Robust Grasp Success

Robust grasp success (%), reproduced from Table III. Higher is better.
Method	Nominal	Wide	Bimodal
Particle-Filter MPC	67%	39%	28%
Gaussian MPC	75%	63%	42%
Gaussian-CVaR MPC	75%	63%	42%
CEM MPC	63%	67%	33%
VNB-MPC (ours)	79%	67%	38%

Wall-Clock Time per Episode

Mean wall-clock time per episode, reproduced from Table III. Lower is better.
Method	Nominal	Wide	Bimodal
Particle-Filter MPC	49.6 s	32.6 s	41.7 s
Gaussian MPC	3.2 s	3.4 s	4.3 s
Gaussian-CVaR MPC	3.1 s	3.4 s	4.3 s
CEM MPC	5.5 s	5.4 s	5.1 s
VNB-MPC (ours)	8.5 s	7.1 s	6.8 s

VNB-MPC achieved higher robust success than particle-filter MPC in all three friction regimes and required 6.8–8.5 s per episode, compared with 32.6–49.6 s. The Gaussian MPC variants were faster; under bimodal friction, they achieved 42% robust success, compared with 38% for VNB-MPC.

Hardware under Pose Uncertainty. Across three objects and four pose offsets (12 trials per method), both VNB and Gaussian MPC achieved 100% grasp-and-lift success. VNB required a median of 6 steps and 11.5 s, compared with 7 steps and 14.2 s for Gaussian MPC (Table IV). This small trial set supports an execution and convergence comparison.

Read the Evaluation in the Paper

Code, Simulation Assets, and Grasp Dataset

Code and Reproduction

Belief models, grasp-quality metrics, experiment runners, and instructions for reproducing the simulation tables.

Reproduction Guide

Simulation Assets

MuJoCo scenes and meshes for the ZArm 622 arm and RealHand L6 hand, with control and pregrasp-planning utilities.

Scenes Meshes and Textures

243 Force-Closed Grasps

GraspIt!-generated grasp candidates stored as JSON databases for the benchmark objects.

Browse the Dataset

BibTeX

@article{enweremVariationalNeuralBeliefParameterizations2026a,
  author = {Enwerem, Clinton and Kalyanaraman, Shreya
            and Baras, John S. and Belta, Calin},
  title = {Variational Neural Belief Parameterizations for
           Robust Dexterous Grasping under Multimodal Uncertainty},
  year = {2026},
  eprint = {2604.25897},
  eprinttype = {arxiv},
  note = {Preprint, arXiv:2604.25897}
}