Learning humanoid reaching tasks in dynamic environments

A central challenging problem in humanoid robotics is to plan and execute dynamic tasks in dynamic environments. Given that the environment is known, sampling-based online motion planners are best suited for handling changing environments. However, without learning strategies, each task still has to be planned from scratch, preventing these algorithms from getting closer to realtime performance. This paper proposes a novel learning-based motion planning algorithm for addressing this issue. Our algorithm, called the attractor guided planner (AGP), extends existing motion planners in two simple but important ways. First, it extracts significant attractor points from successful paths in order to reuse them as guiding landmarks during the planning of new similar tasks. Second, it relies on a task comparison metric for deciding when previous solutions should be reused for guiding the planning of new tasks. The task comparison metric takes into account the task specification and as well environment features which are relevant to the query. Several experiments are presented with different humanoid reaching examples in the presence of randomly moving obstacles. Our results show that the AGP greatly improves both the planning time and solution quality, when comparing to traditional sampling-based motion planners.