Robotic Task Success Evaluation Under Multi-modal Non-Parametric Object Pose Uncertainty

TL;DR

This work tackles predicting robotic task success under uncertain 6D object poses by representing both the estimated pose error and the acceptable task error as multi-modal non-parametric distributions. It introduces an offline-online framework: offline pre-computation of the acceptable error space via dynamic simulations, and online integration over the estimated pose error distribution to compute $P( ext{task}| ext{hat}{o})$. The key contributions are the dual multi-modal uncertainty representations and an offline-to-online mapping that enables reliable task execution decisions with reduced re-planning, demonstrated on two mobile manipulation tasks with improved success rates and fewer failures. The approach offers practical impact by enabling robots to decide when to proceed or wait under pose uncertainty, leveraging dynamic simulations to manage online computational costs and broad applicability to tasks requiring pose estimates.

Abstract

Accurate 6D object pose estimation is essential for various robotic tasks. Uncertain pose estimates can lead to task failures; however, a certain degree of error in the pose estimates is often acceptable. Hence, by quantifying errors in the object pose estimate and acceptable errors for task success, robots can make informed decisions. This is a challenging problem as both the object pose uncertainty and acceptable error for the robotic task are often multi-modal and cannot be parameterized with commonly used uni-modal distributions. In this paper, we introduce a framework for evaluating robotic task success under object pose uncertainty, representing both the estimated error space of the object pose and the acceptable error space for task success using multi-modal non-parametric probability distributions. The proposed framework pre-computes the acceptable error space for task success using dynamic simulations and subsequently integrates the pre-computed acceptable error space over the estimated error space of the object pose to predict the likelihood of the task success. We evaluated the proposed framework on two mobile manipulation tasks. Our results show that by representing the estimated and the acceptable error space using multi-modal non-parametric distributions, we achieve higher task success rates and fewer failures.

Figures (8)

• Figure 1: a) Selected tool pose (grasp and base pose) relative to the object frame $o_{}$b)Acceptable error space pre-computed by introducing errors $\epsilon$ in object frame $o_{}$ and determining the probability of task success c)Estimated error space for the object frame estimate $\hat{o}_{}$ obtained using estimated visual object pose distribution d) The probability of task success is calculated by integrating the acceptable error space over the estimated error space.
• Figure 2: a) Selected base pose $\mathbf{T}_{b}^{o_{}}$ and top-down view of the grasping scene. Error space $\mathcal{E}$ (green box) is defined around the selected base pose for grasping (blue arrow at the center of robot base) b)Acceptable error space$\mathcal{E}_{\text{acc}}^{\hat{o}_{}}$.
• Figure 3: Calculating the $\mathrm{P}(\text{IK}|\hat{o}_{})$a)Estimated error space$\mathcal{E}_{\text{est}}^{\hat{o}_{}}$ (yellow ellipsoid) mapped onto the acceptable error space$\mathcal{E}_{\text{acc}}^{\hat{o}_{}}$. b)Estimated error space$\mathcal{E}_{\text{est}}^{\hat{o}_{}}$ (orange cone) mapped onto the acceptable error space$\mathcal{E}_{\text{acc}}^{\hat{o}_{}}$.
• Figure 4: Calculating the probability of grasp success for the bowl object. a) Selected grasp pose and Acceptable error space$\mathcal{E}_{\text{acc}}^{\hat{o}_{}}$b)Estimated error space$\mathcal{E}_{\text{est}}^{\hat{o}_{}}$c)$\mathcal{E}_{\text{acc}}^{\hat{o}_{}}$ (green) integrated over $\mathcal{E}_{\text{est}}^{\hat{o}_{}}$ (orange) for calculating $\mathrm{P}(\text{grasp}|\hat{o}_{})$.
• Figure 5: Setup for grasping experiments: The default viewpoint and additional available viewpoints. The next viewpoint is selected according to the above order if the robot decides not to perform the task using the pose estimate $\hat{X}_{}$ and distribution $\mathrm{P}(X_{})$ estimated from the current viewpoint.