Multi-resolution Planning for Earthmoving

Research Topic

a multi-resolution planning and execution scheme:

• At the highest level (coarse planner)
  is a coarse planning scheme that uses geometry of the site and the goal configuration of the terrain to plan a sequence of “dig regions”.
• At the lower level (refined planner)
  each dig region is search for the ‘best’ dig that can be executed and finally the selected dig is executed using a force based closed loop control scheme. The control scheme is robust to errors in sensing the geometry of the terrain and is effective in capturing material.

其实就是where to dig $\to$ how to dig

这样做的好处：
The coarse planner ensures even performance over a large number of digs.
The refined planner chooses digs that meet constraints of geometry (reachability and collisions) and optimize a cost function (e.g. volume, energy, time)

原文有有一句话：
“The mechanisms themselves are different and hence the representation used by the coarse and the refined planner is quite different, but at an abstracted level, the process is identical.”
——其实又从侧面印证了本文framework和HRL的相通，i.e. 学习的是skill而不是task.

Multi-resolution Planning

The coarse planner takes as input the terrain map and the location of the truck. The output is a sequence of regions, each of which is in turn sent to a refined planner.
The refined planner operates on an abstract representation of an atomic action, a single dig. Rather than searching for a bucket trajectory, the refined planner searches for the best action within the bounds specified by the coarse planner.

在这个过程中有一个特殊的模块——dig evaluator
the evaluator evaluates candidates using a forward model that simulates the result of choosing an action (在本文中就是the starting location of the bucket)

Coarse Planner

Refined Planner

the refined planner only search for the pose of the bucket as it enters the ground.
the process of excavation is planar, 这是挖掘机不同于常规机械臂的又一个重要特点。
因此可以这样represent an action: two parameters