17 August 2023 in IEEE Robotics and Automation Letters
Actively planning sensor views during object reconstruction is crucial for autonomous mobile robots. An effective method should be able to strike a balance between accuracy and efficiency. In this paper, we propose a seamless integration of the emerging implicit representation with the active reconstruction task. We build an implicit occupancy field as our geometry proxy. While training, the prior object bounding box is utilized as auxiliary information to generate clean and detailed reconstructions. To evaluate view uncertainty, we employ a sampling-based approach that directly extracts entropy from the reconstructed occupancy probability field as our measure of view information gain. This eliminates the need for additional uncertainty maps or learning. Unlike previous methods that compare view uncertainty within a finite set of candidates, we aim to find the next-best-view (NBV) on a continuous manifold. Leveraging the differentiability of the implicit representation, the NBV can be optimized directly by maximizing the view uncertainty using gradient descent. It significantly enhances the method's adaptability to different scenarios. Simulation and real-world experiments demonstrate that our approach effectively improves reconstruction accuracy and efficiency of view planning in active reconstruction tasks. The proposed system is open-sourced at https://github.com/HITSZ-NRSL/ActiveImplicitRecon.git .
26 April 2024 in IEEE/ASME Transactions on Mechatronics
Six-DoF grasp detection has attracted increasing attention due to its high flexibility in manipulation tasks. Recent learning-based methods typically use point sampling or filtering to extract a small subset of scene points for grasp pose computation. However, the sampling method disregards the fact that each scene point contributes differently to the grasp, while the filtering method nearly destroys the geometrical structure of the scene. Concurrently, indirectly learning point features by inputting point coordinates to the network is insufficient for detecting grasp poses since this task heavily depends on the geometrical structure of the local point clouds. In order to solve the above-mentioned problems, we design a novel interest point selection algorithm in the grasp scene, and a novel explicit geometrical feature extraction algorithm for local point clouds in the six-DoF grasp pose detection task. We include the basic grasp points, nearby object points and background points as interest points in calculating the grasp poses. Moreover, we add a module that predicts grasp collision within the network to eliminate the time-consuming collision detection in postprocessing. The proposed network is trained and evaluated on the GraspNet1Billion dataset, and achieves state-of-the-art performance. We also conduct an ablation study to validate the proposed algorithms and the network. Furthermore, we successfully implement our network on a UR5e robotic arm to demonstrate its practical applicability in both single-object and multiobject grasp scenarios.