Workshops

Flexible manipulators are increasingly important in modern robotics because of their lightweight structure, low energy consumption, and ability to perform fast motion. However, the flexibility of the links introduces vibration, tip deflection, uncertainty, and payload-dependent dynamic effects, making accurate trajectory tracking a challenging task. This talk presents a robust control framework for a planar two-link flexible manipulator using a barrier-function-based adaptive second-order non-singular terminal sliding mode control strategy. The proposed approach aims to achieve accurate tip-trajectory tracking while simultaneously suppressing elastic vibration and tip deflection. The adaptive barrier function adjusts the control gain online, avoiding excessive gain estimation and removing the need for prior knowledge of disturbance bounds. The second-order non-singular terminal sliding mode structure improves convergence speed, reduces chattering, and ensures finite-time tracking performance. In addition, the talk discusses payload-dependent modal analysis to understand how different payload conditions influence the natural frequencies and vibration characteristics of the manipulator. Simulation results show that the proposed controller provides improved tracking accuracy, smoother control effort, and better vibration suppression compared with existing sliding mode control approaches. The presentation highlights the significance of robust and adaptive control for flexible robotic systems operating under uncertainty and varying payload conditions.

Ensemble simulations used for forecasting extreme weather events generate increasingly large and complex output data due to dramatic improvements in computational performance. Consequently, it has become proportionately more difficult to efficiently extract noteworthy features and anomalies from such a huge amount of data, as well as to analyze the behavior of individual members and the detailed relationships among variables. In this study, we developed a visual analysis system that integrates an Angular-based edge bundled Parallel Coordinates Plot, which achieves optimal axis reordering to overview correlations across all members, with a spatial analysis view targeting the two variables of interest extracted from this process, and validated its effectiveness.

Large-scale, high-resolution numerical simulations are a critical foundation for scientific discovery. However, analytical workflows that seek to understand phenomena by varying simulation conditions, tracking temporal evolution, and observing results from arbitrary viewpoints require repeated simulation runs and visualization with large-scale data I/O, which severely hinders the exploration cycle. To address this challenge, we propose VisGS, a visualization surrogate model based on 3D Gaussian Splatting (3DGS). By directly inferring rendered images from conditioning inputs such as simulation parameters, timesteps, and viewpoints, VisGS replaces at inference time the conventional end-to-end pipeline from simulation execution to rendering. As a result, within the condition space covered during training, VisGS enables interactive and exploratory visualization of dynamic scene changes from arbitrary viewpoints without re-running expensive numerical simulations or incurring large-scale data I/O. To adapt the model to scientific visualization data, we introduce a 3D Gaussian initialization pipeline that leverages the underlying numerical data, a deformation network for multidimensional conditioning variables, and a loss function that stabilizes learning under abrupt image changes caused by the emergence and disappearance of phenomena. Experimental results show that the proposed method achieves a rendering speedup of more than 30× over ViSNeRF, a representative NeRF-based method, while maintaining comparable rendering quality. These results demonstrate that VisGS enables real-time exploration and analysis across temporal evolution, viewpoints, and parameter settings for large-scale simulation results.

We introduce the Polygonal Proximity Map (PPM), a compact map that implicitly describes convex free space defined by proximity points. We also present a path planning algorithm, A*QP which exploits the geometric properties of the PPM. By combining these techniques, spatial information of free space can be represented with lightweight data while enabling efficient path planning. Experimental results demonstrate that the proposed approach achieves superior performance in terms of computation efficiency and data size compared with occupancy grid maps.

This talk presents “MATT-Diff”, a multi-modal control policy for active multi-target tracking in partially observable environments. By training a diffusion policy to imitate diverse expert strategies ranging from frontier exploration to RRT*-based tracking, our approach effectively balances the dilemma between exploring unseen areas to detect new targets and tracking detected targets. Experimental results demonstrate that MATT-Diff achieves superior zero-shot generalization in unseen environments of out-of-disribution, and outperforms traditional learning-based baselines in reducing target uncertainty.