DART Lab

Robotic Spatial Augmented Reality (RSAR) systems present a unique control challenge as their end-effector is a projection, whose final position depends on both the actuator’s pose and the external environment’s geometry. Accurately controlling this projection first requires predicting the 6-DOF pose of a projector-camera unit from joint angles; however, loose kinematic specifications in many RSAR setups make precise analytical models unavailable for this task. This study proposes a novel deep learning model combining Long Short-Term Memory (LSTM) and an Attention Mechanism (LSTM–Attention) to accurately estimate the forward kinematics of a 2-axis Pan-Tilt actuator. To ensure a fair evaluation of intrinsic model performance, a simulation framework using Unity and unified robot description format was developed to generate a noise-free benchmark dataset. The proposed model utilizes a multi-task learning architecture with a geodesic distance loss function to optimize 3-dimensional position and 4-dimensional quaternion rotation separately. Quantitative results show that the proposed LSTM–Attention model achieved the lowest errors (Position MAE: 18.00 mm; Rotation MAE: 3.723 deg), consistently outperforming baseline models like Random Forest by 9.5% and 17.6%, respectively. Qualitative analysis further confirmed its superior stability and outlier suppression. The proposed LSTM–Attention architecture proves to be a effective and accurate methodology for modeling the complex non-linear kinematics of RSAR systems.

Three-dimensional (3D) geospatial technologies are essential in urban digital twins, smart cities, and metaverse. Rendering large-scale terrain data, often exceeding tens of terabytes, presents challenges. While planetary-scale platforms, like Google Earth and Cesium stream data, the streaming of data and the use of regular grid-type digital elevation models lead to cracks among tiles with different levels of detail. This paper proposes a novel dynamic tile-map generation method to eliminate these cracks. Unlike existing methods, our approach leverages tile subindex information to efficiently construct a tile adjacency map, significant reducing the search space for neighboring tiles and eliminating the need for prior knowledge of the terrain tile structure. Furthermore, our approach is robust to data loss, mitigating cracks caused by missing or incomplete tiles. Compared with existing root-down search methods, our method reduces processing time by 1–5 ms per frame and decreases the number of tile-to-tile links by a factor of 3–5, as demonstrated by experimental results.

Deep Reinforcement Learning (DRL) algorithms often exhibit significant performance variability across different training runs, even with identical settings. This paper investigates the hypothesis that a key contributor to this variability is the divergence in the observation spaces explored by individual learning agents. We conducted an empirical study using Proximal Policy Optimization (PPO) agents trained on eight Atari environments. We analyzed the collected agent trajectories by qualitatively visualizing and quantitatively measuring the divergence in their explored observation spaces. Furthermore, we cross-evaluated the learned actor and value networks, measuring the average absolute TD-error, the RMSE of value estimates, and the KL divergence between policies to assess their functional similarity. We also conducted experiments where agents were trained from identical network initializations to isolate the source of this divergence. Our findings reveal a strong correlation: environments with low-performance variance (e.g., Freeway) showed high similarity in explored observation spaces and learned networks across agents. Conversely, environments with high-performance variability (e.g., Boxing, Qbert) demonstrated significant divergence in both explored states and network functionalities. This pattern persisted even when agents started with identical network weights. These results suggest that differences in experiential trajectories, driven by the stochasticity of agent–environment interactions, lead to specialized agent policies and value functions, thereby contributing substantially to the observed inconsistencies in DRL performance.

This paper presents a methodology for correcting geometric and topological errors, specifically addressing fragmented and disconnected components in buildings (FDCB) in 3D models intended for urban digital twin (UDT). The proposed two-stage approach combines geometric refinement via duplicate vertex removal with topological refinement using a novel spatial partitioning-based Depth-First Search (DFS) algorithm for connected mesh clustering. This spatial partitioning-based DFS significantly improves upon traditional graph traversal methods like standard DFS, breadth-first search (BFS), and Union-Find for connectivity analysis. Experimental results demonstrate that the spatial DFS algorithm significantly improves computational speed, achieving processing times approximately seven times faster than standard DFS and 17 times faster than BFS. In addition, the proposed approach achieves a data size ratio of approximately 20% in the simplified mesh, compared to the 50–60% ratios typically observed with established techniques like Quadric Decimation and Vertex Clustering. This research enhances the quality and usability of 3D building models with FDCB issues for UDT applications.