Online Motion Planning for Robot Manipulators in Dynamic Environments

This thesis focuses on the development of motion planning algorithms for dynamic environments. A dynamic environment is defined as one that evolves over time, necessitating that planning algorithms be efficient in terms of planning and replanning and capable of adapting to the robot's current state and to potential changes in the environment.
To address these challenges, this thesis proposes three contributions, each targeting a different aspect of motion planning in dynamic settings.

In order to achieve efficient planning and replanning with short planning time and small variations, the first research question is formulated to address the bottleneck of speeding up motion planning methods. Since collision checking is a major bottleneck for rapid motion planning, a motion planning method is proposed to identify and eliminate unnecessary collision checks during planning.
This method uses a precomputed deterministic roadmap to capture the collision-free space of the static environment. It then performs a heuristic-informed search on the roadmap and introduces a novel safe zones concept for edge examination to find a feasible solution.
... mehr
When exploring the precomputed roadmap for a solution, the heuristic-informed search minimizes the number of edge examinations by prioritizing edges based on heuristics derived from path costs in the roadmap. The concept of safe zones utilizes the spatial relation between the robot and the environment to identify regions that do not require collision checks. As a result, the method achieves both a reduced number of edge examinations and a decrease in collision checks for each examination.

Despite the significant speedup by reducing the number of collision checks, the planning time naturally increases with problem complexity and remains unbounded. The second contribution is a method to generate subgoals for decomposing complex motion planning problems into small, easily solvable subproblems. Iteratively planning the subproblems ensures short planning time in dynamic environments. This method begins with a pipeline to collect a dataset of suitable subgoals, which can be planned within a specific bounded time. Using this dataset, a conditional generative model captures the distribution of subgoals based on the given planning problems. During inference, a time estimator acts as a critic to evaluate whether the generated subgoals can meet the desired time constraints and lead the robot toward the final goal. Based on this evaluation, the proposed method makes informed decisions to select a suitable subgoal for planning.

The contributions above solely consider the current geometric state of dynamic environments and overlook the temporal aspect. This short-sightedness over the time horizon can result in frequent replanning and local-minima issues. The third contribution addresses this issue by using episodic reinforcement learning to implicitly account for potential changes in the environment over time. The learned policy refines reference trajectories based on the dynamic environment.
Three trajectory refinement strategies based on B-spline movement primitives are introduced to modify the reference trajectories while ensuring smooth trajectory transitions. The planning results of the contributions above can be used as reference trajectories. This spatiotemporally aware method achieves superior task performance compared to methods that solely depend on geometric information and other spatiotemporal planners.

Comprehensive evaluations and ablation studies in simulation and real robot experiments are conducted to demonstrate the effectiveness and limitations of the proposed methods.