Mixed-Integer and Conditional Trajectory Planning for an Autonomous Mining Truck in Loading/Dumping Scenarios: A Global Optimization Approach

Abstract

Trajectory planning for a heavy-duty mining truck near the loading/dumping sites of an open-pit mine is difficult. As opposed to trajectory planning for a small-sized passenger car in a parking lot, trajectory planning for a heavy-duty mining truck involves complex factors in vehicle kinematics and environment. These factors make the concerned trajectory planning scheme a mixed-integer nonlinear program (MINLP) incorporated with conditional constraints (denoted as C-MINLP). MINLP solvers can neither deal with conditional constraints nor find global optima in real time. Instead of solving the C-MINLP directly, we build a from-coarse-to-fine framework so that the coupled difficulties (the mixed integral variables, conditional constraints, and the demand for global optimality) are divided and conquered. At the coarse search stage, a global-optimality-enhanced hybrid A* search algorithm is proposed to find a near-optimal coarse trajectory with the mixed integral variables, conditional kinematic constraints, and global optimality considered. The coarse trajectory is further polished at the refinement stage, wherein the nominal C-MINLP is simplified as a small-scale NLP. The solution to the NLP is an optimized trajectory, which does not violate the complex constraints in the nominal C-MINLP. This indicates that conversion from the C-MINLP to an NLP is efficient with the help of a high-quality coarse trajectory.

Hierarchical Interpretable Imitation Learning for End-to-End Autonomous Driving

“…This strategy, denoted as path velocity decomposition (PVD), has been commonly used because it converts a 3D problem into two 2D ones, which largely facilitates the solution process. Conversely, non-PVD methods directly plan trajectories, which has the underlying merit to improve the solution optimality [18], [37]- [39].…”

Online Competition of Trajectory Planning for Automated Parking: Benchmarks, Achievements, Learned Lessons, and Future Perspectives

“…Recall that an on-road trajectory planner is typically developed in the Frenet frame (also known as a curvilinear coordinate system or a road-aligned frame) [14], [15], which makes all the surrounding obstacles located to the left or right side of the ego vehicle. By contrast, parking planners commonly work in the Cartesian frame, wherein obstacles in a cluttered environment render multiple homotopy classes for the parking planner to choose [16], [17]. Herein, an imperfect choice would degrade the parking efficiency or even lead to a solution failure.…”

“…Notably, some well-known planners, such as the hybrid A * (HA) search algorithm [26], only aim to derive a feasible but not optimal homotopy class. Thus, their planning performances may be unstable in dealing with cases with multiple homotopy classes [17].…”

Adaptive Pure Pursuit: A Real-Time Path Planner Using Tracking Controllers to Plan Safe and Kinematically Feasible Paths

“…Li et al [19] noticed that massive obstacles would render massive homotopy classes, thus an optimal homotopy class deserves to be greedily selected in a parallel-computation architecture. Similarly, an optimality-enhanced hybrid A* search algorithm [20] is proposed, which continues to sample better paths even after a feasible one is already available. The output of a sampling-based planner is kinematically feasible and collisionfree, but the derived path is commonly sub-optimal, jerky, and non-smooth [21].…”