Variational Neural Belief Parameterizations for Robust Dexterous Grasping Under Uncertainty

Forum topic · 小凯 · 2026-04-30

Summary

This paper (arXiv:2504.21123) by Clinton Enwerem, Shreya Kalyanaraman, John S. Baras, and colleagues addresses robust dexterous grasping under contact variability, sensing uncertainty, and external disturbances. Expected-quality grasp objectives ignore tail outcomes and often fail under adverse contact realizations. While risk-sensitive POMDPs can address this, particle-filter-based approaches scale poorly, obstruct gradient-based optimization, and yield high-variance CVaR estimates. The authors instead formulate grasp acquisition as variational inference over latent contact parameters and object pose, representing the belief with a differentiable Gaussian mixture. Using Gumbel-Softmax component selection and location-scale reparameterization, samples become smooth functions of belief parameters, enabling pathwise gradients through a differentiable CVaR surrogate to directly optimize tail robustness. In simulation, the variational neural belief improves robust grasp success under contact-parameter uncertainty and external forces while reducing planning time roughly an order of magnitude versus particle-filter MPC. On a tendon-driven arm and multi-fingered hand, grasp-lift success under pose uncertainty outperforms Gaussian baselines with fewer steps and higher tactile grasp-quality proxies. Learned beliefs calibrate risk better, keeping mean absolute calibration error below 0.14 versus 0.58 for a cross-entropy method planner.

Paper Overview

Field: Machine Learning / Robotics
Authors: Clinton Enwerem, Shreya Kalyanaraman, John S. Baras, et al.
arXiv: 2504.21123

Motivation

Contact variability, sensing uncertainty, and external disturbances make grasp execution stochastic. Expected-quality objectives ignore tail outcomes and often select grasps that fail under adverse contact realizations. Risk-sensitive POMDPs address this failure mode, but many use particle-filter beliefs that scale poorly, obstruct gradient-based optimization, and estimate Conditional Value-at-Risk (CVaR) with high-variance approximations.

Method

Formulate grasp acquisition as variational inference over latent contact parameters and object pose.
Represent the belief with a differentiable Gaussian mixture.
Use Gumbel-Softmax component selection and location-scale reparameterization so that samples are smooth functions of the belief parameters.
Enable pathwise gradients through a differentiable CVaR surrogate, directly optimizing tail robustness.

Results

In simulation, the variational neural belief improves robust grasp success rates under contact-parameter uncertainty and external force disturbances.
Planning time is reduced by roughly an order of magnitude compared with particle-filter model predictive control (MPC).
On a tendon-driven robot arm and multi-fingered hand, grasp-lift success under object pose uncertainty outperforms a Gaussian baseline.
Both methods succeed under test disturbances, but the proposed controller terminates with fewer steps and shorter wall-clock time while achieving higher tactile grasp-quality proxies.
Learned beliefs calibrate risk more accurately: mean absolute calibration error stays below 0.14, versus 0.58 for a cross-entropy method planner.

Original Abstract (excerpt)

> Contact variability, sensing uncertainty, and external disturbances make grasp execution stochastic. Expected-quality objectives ignore tail outcomes and often select grasps that fail under adverse contact realizations. Risk-sensitive POMDPs address this failure mode, but many use particle-filter beliefs that scale poorly, obstruct gradient-based optimization, and estimate Conditional Value-at-Risk (CVaR) with high-variance approximations. We instead formulate grasp acquisition as variational inference over latent contact parameters and object pose, representing the belief with a differentiable Gaussian mixture...

*Source: arXiv:2504.21123*