This paper positions autonomous excavators as a new category of intelligent off-road vehicles, extending the perception-planning-control paradigm of autonomous driving to construction environments where machines simultaneously navigate and actively modify their operational workspace. Autonomous excavation requires coordinated planning and control as machines modify terrain under soil variability and kinematic constraints. This paper proposes a hierarchical framework separating task planning, execution parameter optimization, and trajectory generation. The high-level planner combines gradient-boosting trees with Q-learning behavior trees and prioritized experience replay to retain interpretable decisions while generalizing across states. Reinforcement learning determines execution parameters, and imitation learning generates motion trajectories. Combining the planner with learned execution reduces mean absolute height and terrain standard deviation by 37% and 24%, respectively, relative to the planner alone, and by 62% and 78% relative to a rule-based behavior tree baseline. Planner-level decision latency remains below 4 ms. These results support separating auditable task planning from learned execution, and establish autonomous excavators as intelligent construction vehicles that extend intelligent vehicle research to off-road operational domains.
Autonomous multi-UAV systems require coordinated navigation in cluttered environments while maintaining formation geometry and avoiding obstacles and inter-agent conflicts. This problem becomes challenging when global navigation decisions must be coordinated with local avoidance maneuvers, since individual UAVs may need to temporarily deviate from their nominal formation while preserving mission coordination. To address this problem, this paper presents an entropy-regularized discrete actor-critic framework with improved policy learning for intelligent formation navigation and control of multiple unmanned aerial vehicles (UAVs) operating in cluttered 3-dimensional (3D) environments. We propose to apply goal-based heuristic bias to the action logits before assigning probability distributions to guide the agent towards the goal when action-values are nearly equivalent. This improves convergence and navigation efficiency. We also incorporate entropy regularization in the discrete version of the actor-critic algorithm to balance exploration and exploitation, which helps generate effective navigational policies. The trained actor-critic navigational agent helps the leader UAV find a feasible global path, while follower UAVs maintain a desired geometric formation through decentralized feedback control. We also integrate a local reactive obstacle avoidance mechanism to ensure collision-free operation when unexpected obstacles and inter agent proximity threats arise. The formation is dynamically adjusted using local repulsive forces while maintaining global coordination. The results demonstrate improved navigational policies compared with baseline methods, enabling shorter trajectories and efficient navigation. The analysis highlights that the reactive avoidance strategy enhances formation safety and adaptability while maintaining control stability. These findings confirm that the proposed actor-critic based multi-UAV formation system provides an efficient solution for autonomous cooperative navigation in obstacle-cluttered environments.
A three-dimensional multi-aircraft conflict-resolution architecture is developed around vectorized one-step model predictive control (MPC) and an interaction-aware Proximal Policy Optimization (PPO) residual. Closest-point-of-approach (CPA) geometry and protected-volume entry time support horizontal-primary/vertical-fallback supervision. A shared graph actor conditions horizontal corrections on simultaneous MPC actions; an ordered finite-candidate filter checks the smoothed joint avoidance command. Vertical fallback uses MPC-only altitude layering and a CPA-referenced hold. Analysis establishes a conditional sampled execution-safety result, explicit finite-set feasibility conditions, and local stability of the unsaturated deterministic recovery dynamics. In three eight-aircraft scenarios, the proposed method and its no-residual counterpart each achieved 30/30 safe completions. The residual reduced mean maximum cross-track error by 6.8% and 10.4% in S1 and S2; S3 was unchanged under vertical fallback. Recorded 500-episode training diagnostics distinguish shaped return from deterministic checkpoint selection. The residual-on/off pair shares the full control backbone; common-seed benchmark comparisons and component ablations separate safety evidence from route-efficiency gains. These findings apply to the stated kinematics and encounter families.
Autonomous driving systems must operate reliably in safety-critical scenarios, particularly those involving unusual or complex behavior by Vulnerable Road Users (VRUs). Identifying these edge cases in driving datasets is essential for robust evaluation and generalization, but retrieving such rare human behavior scenarios within the long tail of large-scale datasets is challenging. To support targeted evaluation of autonomous driving systems in diverse, human-centered scenarios, we propose a novel context-aware motion retrieval framework. Our method combines Skinned Multi-Person Linear (SMPL)-based motion sequences and corresponding video frames before encoding them into a shared multimodal embedding space aligned with natural language. Our approach enables the scalable retrieval of human behavior and their context through text queries. This work also introduces our dataset WayMoCo, an extension of the Waymo Open Dataset. It contains automatically labeled motion and scene context descriptions derived from generated pseudo-ground-truth SMPL sequences and corresponding image data. Our approach outperforms state-of-the-art models with up to a 27.5% relative improvement in Top-1 accuracy in motion-context retrieval, when evaluated on the WayMoCo dataset.
Technological innovations are revolutionizing the way we perceive and engage with mobility, with modular vehicle technology (MVT) offering substantial transformative potential for transportation and logistics. MVT enables vehicles to couple and decouple while in motion, supporting applications such as public transportation, freight logistics, car sharing, emergency response, and autonomous vehicle fleets, thereby enhancing the efficiency, flexibility, and adaptability of transportation systems. However, despite its recognized benefits, limited research exists on how modular vehicles (MVs) interact with surrounding traffic during the coupling process and the broader implications of these interactions. Consequently, this study proposes dynamic lane-changing strategies incorporating vehicles' cooperative behavior, together with a coupling approach for MVs. It also investigates the influence of traffic on the MV coupling process and the subsequent impact of MV coupling on overall traffic performance. The proposed conservative (S-C) and aggressive (S-A) lane-changing strategies were benchmarked against VISSIM's built-in lane-changing model (S-V) and evaluated using VISSIM under various traffic demands and speed limits. Across the investigated conditions, coupling failure rates ranged from 33–97%, 3–43%, and 0–10% for S-C, S-V, and S-A, respectively. Furthermore, S-A reduced the distance traveled before coupling initiation by approximately 54–90% relative to S-V and 90–94% relative to S-C. The maximum observed increases in average vehicle delay relative to the base scenario, where MVT is not implemented, were 7.34%, 1.80%, and 4.76% for S-V, S-C, and S-A, respectively. Statistical analyses further confirmed the robustness of the reported coupling and traffic performance findings across the investigated operating conditions. This study provides practical insights for researchers and policymakers and highlights the importance of developing innovative solutions to enable reliable modular vehicle coupling in mixed-traffic environments—an essential step toward real-world deployment.
