Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Finding safe 3D robot grasps through efficient haptic exploration with unscented Bayesian optimization and collision penalty

Joao Castanheira, Pedro Vicente, Ruben Martinez-Cantin, Lorenzo Jamone, Alexandre Bernardino

TL;DR

The paper tackles robust 3D robotic grasping under input uncertainty by extending unscented Bayesian optimization (UBO) to 3D haptic exploration and introducing a collision penalty (CP). Using a Gaussian process surrogate and unscented expected improvement, the approach accounts for input noise and avoids unsafe regions, while the Grasp Wrench Volume provides a principled grasp quality metric. Empirical results in Simox with iCub demonstrate that 3D UBO with CP converges faster and yields safer, higher-quality grasps than standard BO, and that UBO generalizes well from 2D to 3D. The work highlights practical gains in sample efficiency and safety, with potential extensions to full 6D pose optimization and real-world 3D-force-sensing hands.

Abstract

Robust grasping is a major, and still unsolved, problem in robotics. Information about the 3D shape of an object can be obtained either from prior knowledge (e.g., accurate models of known objects or approximate models of familiar objects) or real-time sensing (e.g., partial point clouds of unknown objects) and can be used to identify good potential grasps. However, due to modeling and sensing inaccuracies, local exploration is often needed to refine such grasps and successfully apply them in the real world. The recently proposed unscented Bayesian optimization technique can make such exploration safer by selecting grasps that are robust to uncertainty in the input space (e.g., inaccuracies in the grasp execution). Extending our previous work on 2D optimization, in this paper we propose a 3D haptic exploration strategy that combines unscented Bayesian optimization with a novel collision penalty heuristic to find safe grasps in a very efficient way: while by augmenting the search-space to 3D we are able to find better grasps, the collision penalty heuristic allows us to do so without increasing the number of exploration steps.

Finding safe 3D robot grasps through efficient haptic exploration with unscented Bayesian optimization and collision penalty

TL;DR

The paper tackles robust 3D robotic grasping under input uncertainty by extending unscented Bayesian optimization (UBO) to 3D haptic exploration and introducing a collision penalty (CP). Using a Gaussian process surrogate and unscented expected improvement, the approach accounts for input noise and avoids unsafe regions, while the Grasp Wrench Volume provides a principled grasp quality metric. Empirical results in Simox with iCub demonstrate that 3D UBO with CP converges faster and yields safer, higher-quality grasps than standard BO, and that UBO generalizes well from 2D to 3D. The work highlights practical gains in sample efficiency and safety, with potential extensions to full 6D pose optimization and real-world 3D-force-sensing hands.

Abstract

Robust grasping is a major, and still unsolved, problem in robotics. Information about the 3D shape of an object can be obtained either from prior knowledge (e.g., accurate models of known objects or approximate models of familiar objects) or real-time sensing (e.g., partial point clouds of unknown objects) and can be used to identify good potential grasps. However, due to modeling and sensing inaccuracies, local exploration is often needed to refine such grasps and successfully apply them in the real world. The recently proposed unscented Bayesian optimization technique can make such exploration safer by selecting grasps that are robust to uncertainty in the input space (e.g., inaccuracies in the grasp execution). Extending our previous work on 2D optimization, in this paper we propose a 3D haptic exploration strategy that combines unscented Bayesian optimization with a novel collision penalty heuristic to find safe grasps in a very efficient way: while by augmenting the search-space to 3D we are able to find better grasps, the collision penalty heuristic allows us to do so without increasing the number of exploration steps.
Paper Structure (14 sections, 9 equations, 5 figures, 2 tables)

This paper contains 14 sections, 9 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Bayesian optimization
  3. Surrogate model estimations
  4. Decision using acquisition function
  5. Unscented Bayesian optimization
  6. Unscented expected improvement
  7. Unscented optimal incumbent
  8. Grasp quality metric and collision penalty
  9. Implementation
  10. Experimental results
  11. Benefits of CP
  12. Generalization to 3D
  13. Advantages of UBO over BO
  14. Conclusions

Figures (5)

  • Figure 1: Examples of objects to perform grasp optimization on simulation. Initial pose for each test object.
  • Figure 2: CP vs $\overline{\text{CP}}$ (UBO 3D). Left: expected outcome of current optimum $\Bar{\text{y}}_{\text{mc}}({}\mathbf{x}^{\text{ub}\textit{opt}}_n)$, Right: Variability of the outcome $\text{std}(\text{y}_{\text{mc}}({}\mathbf{x}^{\text{ub}\textit{opt}}_n))$. Best seen in color.
  • Figure 3: 2D vs 3D (UBO CP). Left: expected outcome of current optimum $\Bar{\text{y}}_{\text{mc}}({}\mathbf{x}^{\text{ub}\textit{opt}}_n)$, Right: Variability of the outcome $\text{std}(\text{y}_{\text{mc}}({}\mathbf{x}^{\text{ub}\textit{opt}}_n))$. Best seen in color.
  • Figure 4: BO vs UBO (3D CP). Left: expected outcome of current optimum $\Bar{\text{y}}_{\text{mc}}(\mathbf{x}^\textit{opt}_n)$, Right: Variability of the outcome $\text{std}(\text{y}_{\text{mc}}(\mathbf{x}^\textit{opt}_n))$. Best seen in color.
  • Figure 5: Best grasps in one of the runs. The best grasp of UBO 2D CP (a) is similar to the UBO 3D CP (c). The best grasp achieved by BO is in an unsafe zone (b). The UBO's best grasp is more robust to input noise (c). Check Table \ref{['tab:bestgrasps']} for the grasp metrics in these configurations.