8.1 Motivation

For most problems involving the physical world, some form of feedback is needed. This means the actions of a plan should depend in some way on information gathered during execution. The need for feedback arises from the unpredictability of future states. In this chapter, every state space will be either discrete, or $ X = {\cal C}$, which is a configuration space as considered in Chapter 4.

Two general ways to model uncertainty in the predictability of future states are

  1. Explicitly: Develop models that explicitly account for the possible ways that the actual future state can drift away from the planned future state. A planning algorithm must take this uncertainty directly into account. Such explicit models of uncertainty are introduced and incorporated into the planning model in Part III.
  2. Implicitly: The model of state transitions indicates that no uncertainty is possible; however, a feedback plan is constructed to ensure that it knows which action to apply, just in case it happens to be in some unexpected state during execution. This approach is taken in this chapter.
The implicit way to handle this uncertainty may seem strange at first; therefore, some explanation is required. It will be seen in Part III that explicitly modeling uncertainty is extremely challenging and complicated. The requirements for expressing reliable models are much stronger; the complexity of the problem increases, making algorithm design more difficult and leading to greater opportunities to make modeling errors. The implicit way of handling uncertainty in predictability arose in control theory [108,122,686]. It is well known that a feedback control law is needed to obtain reliable performance, yet it is peculiar that the formulation of dynamics used in most contexts does not explicitly account for this. Classical control theory has always assumed that feedback is crucial; however, only in modern branches of the field, such as stochastic control and robust control, does this uncertainty get explicitly modeled. Thus, there is a widely accepted and successful practice of designing feedback control laws that use state feedback to implicitly account for the fact that future states may be unpredictable. Given the widespread success of this control approach across numerous applications over the past century, it seems valuable to utilize this philosophy in the context of motion planning as well (if you still do not like it, then jump to Chapter 10).

Figure 8.1: By separating the issue of dynamics from feedback, two less-investigated topics emerge.
% latex2html id marker 40349
...} &
Traditional control theory \ \hline

Due to historical reasons in the development of feedback control, it often seems that feedback and dynamics are inseparable. This is mainly because control theory was developed to reliably alter the behavior of dynamical systems. In traditional motion planning, neither feedback nor dynamics is considered. A solution path is considered open loop, which means there is no feedback of information during execution to close the loop. Dynamics are also not handled because the additional complications of differential constraints and higher dimensional phase spaces arise (see Part IV).

By casting history aside and separating feedback from dynamics, four separate topics can be made, as shown in Figure 8.1. The topic of open-loop planning that involves dynamics has received increasing attention in recent years. This is the focus throughout most of Part IV. Those fond of classical control theory may criticize it for failing to account for feedback; however, such open-loop trajectories (paths in a phase space) are quite useful in applications that involve simulations. Furthermore, a trajectory that accounts for dynamics is more worthwhile in a decoupled approach than using a path that ignores dynamics, which has been an acceptable practice for decades. These issues will be elaborated upon further in Part IV.

The other interesting topic that emerges in Figure 8.1 is to develop feedback plans for problems in which there are no explicit models of dynamics or other differential constraints. If it was reasonable to solve problems in classical motion planning by ignoring differential constraints, one should certainly feel no less guilty designing feedback motion plans that still neglect differential constraints.8.1 This uses the implicit model of uncertainty in predictability without altering any of the other assumptions previously applied in traditional motion planning.

Even if there are no unpredictability issues, another important use of feedback plans is for problems in which the initial state is not known. A feedback plan indicates what action to take from every state. Therefore, the specification of an initial condition is not important. The analog of this in graph algorithms is the single-destination shortest-path problem, which indicates how to arrive at a particular vertex optimally from any other vertex. Due to this connection, the next section presents feedback concepts for discrete state spaces, before extending the ideas to continuous spaces, which are needed for motion planning.

For these reasons, feedback motion planning is considered in this chapter. As a module in a decoupled approach used in robotics, feedback motion plans are at least as useful as a path computed by the previous techniques. We expect feedback solutions to be more reliable in general, when used in the place of open-loop paths computed by traditional motion planning algorithms.

Steven M LaValle 2012-04-20