Learning-based Motion Planning for a Timber Crane

Motion planning is an essential part of robotics research, requiring algorithms that are computationally efficient and adaptable to varying environments. Using machine learning methods for motion planning can provide solutions to these challenges. In this thesis, imitation learning methods for motion planning are investigated, utilizing artificial neural networks to implement these techniques. The networks generate trajectories by imitating two algorithms: Via-point-based Stochastic Trajectory Optimization and the solution to an optimal control problem. These learning-based motion planning methods are applied to a timber crane across different environmental settings. The results demonstrate that the networks are generally able to learn the underlying task through behavioral cloning and adapt to varying obstacle heights. They also show a significant advantage regarding computational speed over the original algorithms. However, in the more complex scenario with two movable obstacles, further improvements are required.