Robots in the real world need to manipulate kinematic mechanisms like doors, drawers, dish washers, oven doors, etc. These mechanisms may have kinematic dependencies that robots need to know and learn about. For example, a door can be locked in its motion by a key and a handle. It's important that robots can effectively and efficiently explore and actuate such mechanisms.
Description of Work
In this work we developed and tested rule based heuristics and more sophisticated planning methods to explore and manipulate complex kinematic mechanisms in a simulation. We compared these different approaches and investigated the effect of different algorithmic features on the exploration procedure. We further analyzed the properties of the environment to clarify which environmental features have an influence on the performance of the algorithms.
We found that dedicated planning for exploration outperforms heuristics. The additional look-ahead and anticipation of potential results of actions enables the agent to explore in a more efficient way, although under certain environmental features heuristics can exhibit similar performance as more costly planning.