Overview
- Field: Machine Learning / Robotics
- Authors: Clinton Enwerem, Shreya Kalyanaraman, John S. Baras, et al.
- Released: 2026-04-29
- arXiv: 2504.21123
- Problem: Grasp execution is stochastic due to contact variability, sensor noise, and external disturbances. Expected-quality objectives ignore tail outcomes and often select grasps that fail under adverse contact realizations.
- Limitation of prior work: Risk-sensitive POMDPs address this failure mode, but many rely on particle-filter beliefs that scale poorly, block gradient-based optimization, and yield high-variance approximations of Conditional Value-at-Risk (CVaR).
- Approach: The authors reformulate grasp acquisition as variational inference over latent contact parameters and object pose. The belief is represented by a differentiable Gaussian mixture.
- Reparameterization tricks: Gumbel-Softmax component selection and location-scale reparameterization express samples as smooth functions of belief parameters.
- Optimization: A differentiable CVaR surrogate allows pathwise gradients, enabling direct optimization of tail robustness.
- Simulation results: The variational neural belief improves robust grasp success under contact-parameter uncertainty and external-force disturbances, while reducing planning time by roughly an order of magnitude versus particle-filter model-predictive control.
- Hardware validation: On a serial-chain arm with a multi-finger hand, the proposed controller matches a Gaussian baseline in success under object-pose uncertainty, but terminates in fewer steps with shorter wall-clock time and yields higher tactile grasp-quality proxies.
- Calibration: The learned belief is better calibrated; mean absolute calibration error stays below 0.14, compared to 0.58 for a cross-entropy-method planner.
Key Points
Source Excerpt (English Abstract)
> 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. We use Gumbel-Softmax component selection and location-scale reparameterization to express samples as smooth functions of the belief parameters, enabling pathwise gradients through...