Task-oriented dexterous grasping remains challenging in robotic manipulations of open-world objects under severe partial observation, where significant missing data invalidates generic shape completion. In this paper, to overcome this limitation, we study \emph{Task-Oriented Shape Completion}, a new task that focuses on completing the potential contact regions rather than the entire shape. We argue that shape completion for grasping should be explicitly guided by the downstream manipulation task. To achieve this, we first generate multiple task-oriented shape completion candidates by leveraging the zero-shot capabilities of object functional understanding from several pre-trained foundation models. A 3D discriminative autoencoder is then proposed to evaluate the plausibility of each generated candidate and optimize the most plausible one from a global perspective. A conditional flow-matching model named FlowGrasp is developed to generate task-oriented dexterous grasps from the optimized shape. Our method achieves state-of-the-art performance in task-oriented dexterous grasping and task-oriented shape completion, improving the Grasp Displacement and the Chamfer Distance over the state-of-the-art by 16.17% and 55.26%, respectively. In particular, it shows good capabilities in grasping objects with severe missing data. It also demonstrates good generality in handling open-set categories and tasks.

The inevitable trade-off between perceptual performance and communication bandwidth in collaborative perception poses a significant challenge. To mitigate this constraint, we introduceHAFNet, a hybrid-stage collaborative perception method designed for multi-agent collaborative 3D object detection. This method initially generates dense proposal, serving as the Region of Interest (RoI). Next, all proposals are aggregated through the proposed Agent-Foreground List via RoI association. Moreover, secondary sampling is performed according to those foreground regions. Furthermore, we achieve superior feature extraction through geometric and offset encoding. Concurrently, the setting of proxy points effectively reduces the size of the collective perception messages. In the end, those features are fused and interacted to get the detection. Extensive experiments on existing DAIR-V2X and V2V4Real datasets illustrate thatHAFNetalmost keeps a minimum communication bandwidth, while it surpasses existing state-of-the-art methods in 3D object detection tasks.

The indeterminacy of human motion poses challenges for pedestrian trajectory prediction. Consequently, existing methods adopt multimodal strategy to model pedestrians future trajectories. A significant advancement in this regard is the growing prominence of the diffusion model. However, the two-dimensional inputs for trajectory prediction not provide sufficient contextual information for the diffusion model. Furthermore, the diffusion model suffers from substantial inference time. To address these conundrums, we propose a trajectory prediction method based on the diffusion model, named as Motion Latent Diffusion (MLD). The core of MLD is the Conditional Variational Autoencoder (CVAE) to transform the original low-dimensional inputs into a higher-dimensional latent space, expanding the receptive field to yield more comprehensive and intricate representations. Simultaneously, during the inferential stage of the diffusion model, we adopt a leapfrogging inference strategy, which facilitates a faster sampling process. Experiments conducted on the ETH/UCY and Stanford Drone datasets (SDD) corroborate the superiority of our method.

Although collaborative 3D perception has successfully improved detection performance by sharing LIDAR information among multiple agents, its impact under adverse weather is under poor investigation. It is non-trivial to reduce the noise effect in the multi-agent system, as each agent may generate defective feature representations with aleatoric uncertainty, and such uncertainty will be further amplified in the collaborative stage due to deterministic collaboration models. To mitigate the negative effects of weather noise on the collaborative framework, we proposed a method called Co-Denoising, which incorporates a two-stage denoising approach within the intermediate collaborative framework. In our method, a sampling-based noise filtering is first performed at each agent to make a coarse denoising. Then, during the collaboration stage, the global feature representations are expanded through Bayesian neural networks to improve the robustness against environmental noise. The extensive experiments on sunny and rainy datasets have indicated the proposed collaborative perception method can significantly reduce performance degradation under adverse weather.

Multi-modal fusion is a promising approach to boost the autonomous driving performance and has already received a large amount of attention. Meanwhile, to increase driving reliability under distinct scenarios, it is important to handle unforeseen weather events in the training dataset, which is known as an Out-Of-Distribution (OOD) problem, for autonomous driving algorithms. In this paper, we consider those two aspects and propose an end-to-end multi-modal domain-enhanced framework, namely CrossFuser, to meet the safety orientated driving requirements. CrossFuser first integrates both image and lidar modalities to generate a robust environmental representation through conjoint mapping, elastic disentanglement, and attention mechanism. Further, the perception embedding is used to calculate corresponding waypoints by a waypoint prediction network, consisting of Gate Recurrent Units (GRUs). Finally, the final control commands are calculated by low-level control functions. We conduct experiments on the Car Learning to Act (CARLA) driving simulator involving complex weather conditions under urban scenarios, the results show that CrossFuser can outperform the state of the art.