Large scale robotic material handling: Learning, planning, and control

本站添加: 2026-05-16

[00:00:04]In this work, we present a comprehensive framework for the autonomous execution of large-scale material handling tasks.

[00:00:12]The transfer of bulk material piles is typically performed by heavy hydraulic manipulators known as material handlers.

[00:00:19]Our framework features four main components to automate bulk material handling tasks, which are an RL-based waypoint following controller, an RL-based throwing controller, an RL-based attack point planner, and an RRT* based path planner.

[00:00:34]We first evaluate our framework on the pile transfer task.

[00:00:40]The RL-based attack point planner processes noisy soil grid heights to determine the next grasping point. After grasping, the unconstrained RL throwing policy throws the material at the designated target location. The cycle then repeats continuously.

[00:00:58]Our framework demonstrates reliable performance over extended operation, successfully relocating all the soil from the original pile.

[00:01:07]Our attack point planner matches human performance in scooped soil per cycle.

[00:01:11]Compared to a human expert operator, our throwing controller achieves more consistent throws to the target location, resulting in a more compact pile with reduced spread and a higher peak. Our framework can adapt to unseen, unstructured work sites with a waypoint planner.

[00:01:26]To evaluate this capability, we create a virtual obstacle in the form of a wall.

[00:01:30]The RRT* planner computes a path before every transition that clears the obstacle with sufficient margin to account for the under-actuated tool.

[00:01:42]The combination of precise throwing and obstacle-aware path planning enables our framework to address a more complex material handling task, dump truck loading.

[00:01:52]To dampen tool oscillations before releasing the material, the expert operator reduces the machine's rotation speed by 41% and significantly slows down the arm movement.

[00:02:02]On the other hand, our framework, designed for such scenarios, reduces the speed by only 24%, bringing the efficiency of our system much closer to the human operator's level. Overall, the framework achieves a performance level comparable to that of a moderately experienced human operator for the dump truck loading task.