7 April 2022 in IEEE Transactions on Robotics
Legged robots are very promising for use in real-world applications, but their operation in narrow spaces is still challenging. One solution for enhancing their environmental adaptability is to design a small-sized biomimetic robot capable of performing multiple motions. By capturing a decent representation of an actual rat ( rattus norvegicus ), we developed a small-sized quadruped robotic rat (SQuRo), which includes four limbs and one flexible spine. On the basis of the extracted key movement joints, SQuRo was subtly designed with a relatively elongated slim body (aspect ratio: 3.42) and smaller weight (220 g) compared with quadruped robots of the same scale. Accordingly, we propose a control framework for multimodal motion planning, and the appropriate control parameters were tuned through optimization with consideration to the stability and actuation limits. The results obtained through a series of experimental tests reveal that SQuRo achieves a superior motion performance compared with existing state-of-the-art small-sized quadruped robots. Remarkably, SQuRo has an extremely small turning radius (0.48 BL) and strong payload capacity (200 g), and it can recover from falls.
22 February 2022 in IEEE Robotics and Automation Letters
Collision-free trajectory planning is a critical technique for space robot mission. In this letter, we developed a model-free Hierarchical Decoupling Optimization (HDO) algorithm to realize 6D-pose multi-target trajectory planning for the free-floating space robot. In order to reduce the complexity of exploration, the whole system consists of two layers: the high-level policy completes the collision-free trajectory planning of the end-effector’s pose; the low-level policy divides the task of reaching arbitrary pose into two decoupling sub-tasks (position and orientation) within a large target space. By introducing the Hindsight Experience Replay (HER), we successfully trained two agents based on multi-goal reinforcement learning. We proposed an Event-based Alternating Optimization (EAO) to stabilize the training and efficiently approximate the optimal policy. Theoretical analysis shows EAO can guarantee the learning stability and reachability of the equilibrium point. The simulation results illustrate that the proposed algorithm achieves high environmental adaptability and anti-disturbance capacity. Furthermore, we demonstrated our proposed method in a practical space mission by applying it to capture a target satellite. Qualitative results (videos) are available at 1.
4 November 2022 in Aerospace Science and Technology
11 December 2025 in IEEE Transactions on Instrumentation and Measurement
Accurate LiDAR-inertial SLAM in dynamic environments remains challenging due to transient moving objects and long-horizon drift accumulation. Point-only formulations offer weak structural constraints, while post-hoc dynamic filtering cannot promptly inform the frontend pose estimation and data association. To address these issues, we present SPLIN, a structured plane-based LiDAR-inertial SLAM framework that couples real-time dynamic suppression with uncertainty-consistent estimation. First, we introduce Incremental Static-referenced Dynamic Object Removal (ISDOR), a lightweight module that detects and suppresses moving objects upon their first appearance at the frame level. Second, we develop a real-time, tightly coupled Point-Plane LiDAR-Inertial Odometry (PPLIO) based on an iterated Kalman filter. PPLIO integrates ISDOR, an efficient plane extraction module, and an adaptive downsampling strategy, enabling robust operation across scenes with widely varying scales and dynamics. To mitigate long-term odometric drift, we further employ uncertainty-aware factor-graph optimization in the backend, yielding a complete, accurate, and efficient LiDAR-inertial SLAM system. Extensive experiments demonstrate the effectiveness of SPLIN: ISDOR yields, on average, a 13.47% F1 improvement (best F1 on 15/17 sequences), a 26.67% precision gain, and a 64.26% reduction in runtime over state-of-the-art baselines. Building on ISDOR, SPLIN achieves the lowest localization error on 28/34 sequences—including highly dynamic and low-feature environments—while maintaining real-time inference at 6.94-36 ms per frame. Ablation studies further validate the contributions of ISDOR and the loop closure strategy to the overall system accuracy.