LastThird

Question 1

What are control architectures?

Answer 1

broad patterns for how a robots software interacts with the environment

Question 2

why are control architectures important?

Answer 2

they provide structure and impose contraints on the way robots function

Question 3

reactive control architectures

Answer 3

respond immediately to sensor readings, then discard them, no memory, can be described as a series of conditionals

Question 4

state machines

Answer 4

slightly bend the rules of reactive CAs, create states where each state is a “mode of operation” with its own controllers, can be represented as a directed graph where the nodes are actions and the edges are the information you get

Question 5

deliberative architecture

Answer 5

think carefully about the effects of each action, use planning algorithms to create plans to achieve goal, representation

Question 6

hybrid architectures

Answer 6

combine both deliberative and reactive elements, consists of a reactive part, deliberative part, and some sort of connection between them

Question 7

what is the tricky part of hybrid architecture?

Answer 7

integrating the various control structure types

Question 8

what is the most common form of hybrid architecture

Answer 8

the hierarchical approach: the deliberative level forms a high-level plan to acheive the robot’s goal that are described in abstract action that the reactive level knows how to execute, and the reactive level executes the most recently generated plan

Question 9

what is a configuration space?

Answer 9

allows us to treat the robot as a point which allows for easier planning and visualization of the robots movements

Question 10

C-Space(configuration space)

Answer 10

contains one point for each combination of values for the robots position, orientation, and internal joint positions, a special kind of state space, often a metric space with a reasonable definition of distance between pairs of configurations, describes where the robot is

Question 11

common C-spaces

Answer 11

SO(2) is a special orthogonal group that can rotate but not translate in the plane, SE(2) is a special euclidean group that can rotate and translate in the plane, and SO(3)/SE(3) are the same but in 3D space

Question 12

obstacle configuration

Answer 12

a configuration in which the robot is in collision with something in the environment, possibly itself, denoted by Cobst. everything else is a free configuration, Cfree = C-Cobst

Question 13

difference between configuration space and work space

Answer 13

configuration space describes all the possible ways that a robot can be arranged, while workspace describes all the points in physical space that the robot’s end-effector can reach

Question 14

collision detection in a C-space

Answer 14

check for collisions between points or short segments only, removes need for constructing a complete representation of obstacles

Question 15

motion planning problem

Answer 15

has three inputs: a description of Cfree, a start configuration, and a goal configuration, and an output: a plan that specifies how the robot should reach its goal

Question 16

issue with combinatorial approach

Answer 16

these algorithms are difficult to understand and implement, take exponential time, likely as bad as NPC, mostly due to the fact that C-space obstacles are hard to compute and represent

Question 17

sampling based algorithms

Answer 17

algorithm chooses a sequence of (usually) random samples that are well distributed in the C-space, then using these samples as a guide it constructs a graph in Cfree, trying to connect the start and goal configurations, can change how the samples are used and what kind of graph is constructed

Question 18

Probabilistic roadmap(PRM)

Answer 18

use a collection of samples plus the start/goal as nodes, discard all samples in Cobst, attempt to connect pairs of nodes that within a connection distance of each other, using collision detection for each edge

Question 19

rapidly exploring random tree(RRT)

Answer 19

begin with the start node, choose a random sample, find the nearest tree node(just the sample at first) to that sample and extend an edge from that tree node to the sample as far as the “connection distance”, then repeat. for collisions can either get rid of edge or stop it at obstacle

Question 20

voronoi bias

Answer 20

the fact that as time goes on, the random samples will pull the tree into unexplored regions

Question 21

completeness

Answer 21

the idea that if a solution exists, the algorithm will find it, PRM and RRT are not complete as success depends entirely on the sequence of samples

Question 22

probabilistic completeness

Answer 22

a motion planner is probabilistically complete if the probability of failure goes to zero as the number of samples increases, bith PRM and RRT are this

Question 23

when is motion planning hard

Answer 23

Motion planning can be particularly hard in dynamic environments where objects are moving and time is constantly changing. It can also be challenging when the robot has many degrees of freedom or when the environment is cluttered with obstacles. Additionally, motion planning can be difficult when the robot must satisfy multiple constraints or when the robot’s movements must be optimized according to some criteria.

Question 24

comparison between PRM and RRT

Answer 24

PRMs are mostly useful when multiple instances need to be solved in the same C-space, one environment and set of obstacles but different tasks. Sine RRTs start from the start node, only work when a single instance needs to be solved, but can handle motion constraints that make direct connections difficult.

Question 25

what d0 asymptotically optimal motion planning algorithms try to do

Answer 25

improve the quality of the solution over time, especially regarding removing extra motion in graphs and getting a shorter path to the goal, so optimal cost and a solution

Question 26

what is the most widely used AO?

Answer 26

optimal RRT or RRT*

Question 27

how does RRT* work

Answer 27

Each node tracks its parent and the total distance to the root along the tree. New nodes are added using the usual sampling, nearest neighbor, and extension processes. Change 1: Each node connects to the best nearby parent, determined by the new node’s distance to the root. Not necessarily the node that was nearest neighbor. Change 2: For each new node, check nearby nodes to see if any would be have smaller distances with the new node as their parent. If so, rewire the tree.
Repeat for other nearby nodes