Increasing Productivity for Autonomous Mass Excavation A Thesis Proposal

This research focuses on the problem of increasing productivity for the task of autonomous mass excavation. Autonomous excavation has the benefits of higher productivity, lower labor costs, increased safety, and the ability to work in hazardous environments. Mass excavation involves rapidly loading trucks with soil/rock/ore using a mobile digging machine with a bucket on an arm-like appendage. It is desirable for this operation to proceed very quickly, load the trucks evenly, avoid excessive spillage, and perform safely while operating in a wide variety of possible digging conditions and worksite configurations. Preliminary work has been done on planning the excavator’s motions using a script-based approach, which takes advantage of the fact that the excavator’s motions are very similar for each bucket load, but the kinematic details can change due to changes in digging, dumping, and truck locations. Currently these kinematic details, also known as the script parameters, are computed using inverse kinematics, simple machine models, heuristics, and “magic” numbers which need to be adjusted from time to time. To date, this technique has worked very well on our test site, with autonomous excavator productivity approaching that of a skilled human operator. It is unclear, however, if the current script parameter computation will achieve the same results on different materials and in different worksite configurations, or if the productivity could be increased even more. This research proposes the idea of using information gathered on-line as the excavator performs its task in its current working conditions as a better way of computing the script parameters. Information about an excavator action, i.e. one set of script parameters, and the results of that action, such as the execution time, location of dumped soil, etc., would be recorded. Robot learning techniques, particularly memory-based learning, have been investigated as a method of using this information to compute the excavator’s next action which will better achieve its performance goals. Preliminary results of a simple learning system show it is possible for the excavator to become faster and more accurate with manipulating the soil. It is believed this proposed approach will result in productivity that equals or exceeds human operators while maintaining the ability to be productive in a wide range of working conditions.