Peng Lu

Event cameras asynchronously output low-latency event streams, promising for state estimation in complex conditions. The motion-dependent nature of event cameras presents persistent challenges in achieving robust event feature detection and matching. Recent learning-based approaches have demonstrated superior robustness over traditional handcrafted methods, particularly under aggressive motion and HDR scenarios. This article proposes SuperEIO, a novel framework that leverages a learning-based event-only detector and IMU measurements for event-inertial odometry. Our event-only feature detector employs a convolutional neural network on continuous event streams, while a graph neural network achieves event descriptor matching for loop closure. We accelerate network inference with TensorRT, ensuring low-latency, real-time operation on resource-constrained devices. Extensive evaluations on multiple public benchmarks demonstrate its superior accuracy and robustness compared with the advanced event-based methods. Moreover, we conduct a large-scale real-world experiment on an edge handheld platform to demonstrate long-term effectiveness. Our pipeline is open-sourced to facilitate research in the field: https://github.com/arclab-hku/SuperEIO.

Efficient navigation in unknown environments relies on a compact yet geometrically expressive map that enables optimal path generation. Among various representations, Voronoi diagram offers advantages in structural compactness and global connectivity. However, in partially explored environments, its limited spatial expressiveness degrades planning performance. To address this, we propose the enveloped generalized Voronoi graph (EGVG) and a multihomotopy path-planning algorithm, integrated within a unified navigation framework. EGVG enhances the original Voronoi skeleton by incorporating Euclidean signed distance field (ESDF) contours to envelop obstacles, enriching spatial representation and bridging the gap between the Voronoi skeleton and the optimal solution space. Next, a multihomotopy search is performed on the EGVG, where the planner deactivates key vertices after each iteration, forcing the exploration of multiple homotopy classes and increasing the likelihood of finding globally shortest paths (i.e., solution optimality). Finally, a motion-consistency criterion is applied to select the most efficient path for execution. Experimental results demonstrate that the proposed method achieves up to an 88% improvement in solution optimality over the original generalized Voronoi graph. Compared to roadmap and visibility-graph representations, it reduces planning time and graph size by over an order of magnitude. In unknown navigation tasks, our planner outperforms state-of-the-art methods, achieving shorter travel distances and execution times.

Unmanned Aerial Vehicles (UAVs) and Unmanned Ground Vehicles (UGVs) offer complementary capabilities for long-duration autonomous missions, where UAV recovery by a ground platform can significantly reduce aerial energy consumption. In practical GPS-denied environments, however, such docking-return operations are limited by three coupled difficulties: establishing an initial relative configuration without external localization infrastructure, selecting an energy-efficient docking point under heterogeneous vehicle constraints, and maintaining feasibility when unknown obstacles are encountered. This paper proposes a relative-initialization and online-replanning framework for cooperative UAV–UGV docking-return. A lightweight front-end initialization module is first developed using a minimal three-node UWB configuration, onboard IMU/odometry increments, and a short excitation trajectory to recover a consistent UAV–UGV-base relative geometry. The initialized configuration is then used in an energy-time docking-point planner that incorporates UAV flight energy, UGV Dubins-type kinematic constraints, asynchronous arrival coordination, Field-Of-View (FOV) feasibility, and obstacle avoidance. The proposed method is positioned as a task-level docking-point planning and replanning framework, rather than a complete low-level docking controller or a full localization system. The relative-initialization module is evaluated in simulation, while the docking-point planning and online replanning layers are validated through Monte Carlo simulations, Gazebo tests, and real-world experiments with controlled initial relative configurations. These results demonstrate the effectiveness of the proposed task-level planner in improving energy-time performance, reducing UAV waiting, and maintaining feasible docking-return execution under the considered cluttered GPS-denied settings.

Addressing the challenge of quadrupedal terrain traversal in complex environments with constrained obstacles, this work presents an integrated planning-control framework synergizing trajectory optimization (TO), nonlinear model predictive control (NMPC), and whole-body control (WBC). The core theoretical contribution of this work is to propose an enhanced multiple-shooting (MS) based constrained differential dynamic programming (DDP) solver for kinodynamic NMPC. A novel derivation is proposed that accounts for the defect between shooting segments during the DDP backward pass, which improves traditional single-shooting (SS) based DDP robustness to initial guesses. Furthermore, the NMPC is integrated with high-level TO and low-level WBC to form a unified optimal trajectory planning and control framework, enabling quadruped robots to execute high-speed dynamic motions with extended flight phases through constrained obstacles, which constitutes the main technical contribution of this research. Experimental validation via high-fidelity simulations and hardware implementation on our self-developed platform ArcDog confirms both the solver’s superiority and the efficacy of the framework.

Perception‐based Deep Reinforcement Learning (DRL) controllers demonstrate impressive performance on challenging terrains. However, existing controllers still face core limitations, struggling to achieve both terrain generality and platform transferability, and are constrained by high computational overhead and sensitivity to sensor noise. To address these challenges fundamentally, we propose a generalized control framework: Mastering a Generalized Contrastive Depth Model (MGDP). We leverage NVIDIA Warp to enable efficient parallel computation of depth images, thereby mitigating the inherent high computational cost. MGDP extracts low‐dimensional terrain feature representations from multi‐modal inputs (depth images and height maps) and integrates an explicit depth map denoising mechanism. This process not only facilitates effective decoupling of perception from dynamics but also significantly reduces the memory. Furthermore, we design terrain‐adaptive reward functions that modulate penalty strengths according to terrain characteristics, enabling the policy to acquire complex locomotion skills (e.g., climbing, jumping, crawling, squeezing) in a single training stage without relying on distillation. Experimental results demonstrate that MGDP not only endows the policy with superior cross‐terrain generalization capability but also enables fast and efficient fine‐tuning across diverse quadruped robot morphologies via its pre‐trained, dynamics‐decoupled perception model. This vigorously advances the development of unified, efficient, and generalized frameworks for quadrupedal locomotion control.

Agile flight through collapsed buildings primarily demands robust and independent control of attitude and position under unpredictable environmental disturbances and complex aerodynamic interactions. Prior studies improve underactuated quadrotors by adding actuators to enable independent pitch or roll control, but introduce control complexity and redundant mechanical load. To overcome these limitations, we achieve five-dimensional (x, y, z, pitch, and yaw) agile flight through a unified tilt mechanism codesigned with an adaptive physics-informed neural network (PINN)-based disturbance rejection controller for aggressive and precise maneuvers. The developed methodology achieves analytical pitch-axis decoupling and incorporates differential flatness-based trajectory planning to guarantee precise and agile flight despite actuator coupling and environmental disturbances. Furthermore, a robust PINN observer enables real-time estimation of external and internal torques, significantly enhancing disturbance rejection in the presence of model uncertainties. Experimental results demonstrate that the proposed approach achieves high tracking accuracy and robust performance under aggressive maneuvers, successfully traversing a 1.2-m-wide tunnel tilted at 60 ${}^{\circ}$ with an overall position root-mean-square error of 0.26 m, which highlights its strong applicability to inspection tasks in complex environments.

While single-agent legged locomotion has witnessed remarkable progress, individual robots remain fundamentally constrained by physical actuation limits. To transcend these boundaries, we introduce Co-jump, a cooperative task where two quadrupedal robots synchronize to execute jumps far beyond their solo capabilities. We tackle the high-impulse contact dynamics of this task under a decentralized setting, achieving synchronization without explicit communication or pre-specified motion primitives. Our framework leverages Multi-Agent Proximal Policy Optimization (MAPPO) enhanced by a progressive curriculum strategy, which effectively overcomes the sparse-reward exploration challenges inherent in mechanically coupled systems. We demonstrate robust performance in simulation and successful transfer to physical hardware, executing multi-directional jumps onto platforms up to 1.5 m in height. Specifically, one of the robots achieves a foot-end elevation of 1.1 m, which represents a 144% improvement over the 0.45 m jump height of a standalone quadrupedal robot, demonstrating superior vertical performance. Notably, this precise coordination is achieved solely through proprioceptive feedback, establishing a foundation for communication-free collaborative locomotion in constrained environments.