\documentclass{article}
\usepackage{fleqn}
\usepackage{epsf}
\usepackage{aima2e-slides}
\begin{document}
\begin{huge}
\titleslide{Robotics}{Chapter 25}
\sf
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Outline}
Robots, Effectors, and Sensors
Localization and Mapping
Motion Planning
Motor Control
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Mobile Robots}
\centerline{\epsffile{figures/Sojourner.eps}\qquad \epsffile{figures/honda_asimo_robot.eps}}
\centerline{\epsffile{figures/raibert-1leg.eps}\qquad \epsffile{figures/robocup.eps}}
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Manipulators}
\vspace*{0.2in}
\epsfxsize=0.7\textwidth
\fig{figures/stanford-arm.eps}
Configuration of robot specified by 6 numbers\al
$\implies$ 6 \defn{degrees of freedom} (DOF)
6 is the minimum number required to position end-effector arbitrarily.\\
For dynamical systems, add velocity for each DOF.
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Non-holonomic robots}
\vspace*{0.2in}
\epsfxsize=0.4\textwidth
\fig{figures/car-like.eps}
A car has more DOF (3) than controls (2), so is \defn{non-holonomic};\\
cannot generally transition between two infinitesimally close configurations
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Sensors}
\defn{Range finders}: sonar (land, underwater), laser range finder, radar (aircraft),\\
tactile sensors, GPS
\vspace*{0.2in}
\centerline{\epsffile{figures/sick.eps}\qquad\epsffile{figures/robot0124bw.eps}}
\defn{Imaging sensors}: cameras (visual, infrared)\\
\defn{Proprioceptive sensors}: shaft decoders (joints, wheels), inertial sensors,\\
force sensors, torque sensors
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Localization---Where Am I?}
Compute current location and orientation (\defn{pose}) given observations:
\vspace*{0.2in}
\epsfxsize=0.7\textwidth
\fig{figures/robotics-ddn.eps}
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Localization contd.}
\vspace*{0.2in}
\twofig{figures/robotics-pic2.eps}{figures/range-scan-model.eps}
Assume Gaussian noise in motion prediction, sensor range measurements
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Localization contd.}
Can use particle filtering to produce approximate position estimate
\vspace*{0.2in}
\threefig{figures/first.ps}{figures/second.ps}{figures/third.ps}
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Localization contd.}
Can also use \defn{extended Kalman filter} for simple cases:
\vspace*{0.2in}
\epsfxsize=0.75\textwidth
\fig{figures/robotics-pic6.eps}
Assumes that landmarks are {\em identifiable}---otherwise,
posterior is multimodal
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Mapping}
Localization: given map and observed landmarks, update pose distribution
Mapping: given pose and observed landmarks, update map distribution
SLAM: given observed landmarks, update pose and map distribution
Probabilistic formulation of SLAM: \al
add landmark locations $L_1,\ldots,L_k$ to the state vector,\al
proceed as for localization
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Mapping contd.}
\vspace*{0.2in}
\centerline{\epsfysize=0.37\textheight\epsffile{figures/arena010.ps}\qquad\epsfysize=0.37\textheight\epsffile{figures/arena020.ps}}
\centerline{\epsfysize=0.37\textheight\epsffile{figures/arena033.ps}\qquad\epsfysize=0.37\textheight\epsffile{figures/arena034.ps}}
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{3D Mapping example}
\vspace*{0.2in}
\twofig{figures/mine-robot.eps}{figures/mine-data.eps}
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Motion Planning}
Idea: plan in \defn{configuration space} defined by the robot's DOFs
\vspace*{0.2in}
\centerline{\epsffile{figures/armExampleWorkSpace.eps}\qquad\epsffile{figures/armExampleConfSpace.eps}}
Solution is a point trajectory in free C-space
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Configuration space planning}
Basic problem: $\infty^d$ states! Convert to \emph{finite} state space.
\defn{Cell decomposition}:\al
divide up space into simple \defn{cells}, \al
each of which can be traversed ``easily'' (e.g., convex)
\defn{Skeletonization}:\al
identify finite number of easily connected points/lines\al
that form a graph such that any two points are connected\al
by a path on the graph
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Cell decomposition example}
\vspace*{0.2in}
\centerline{\epsffile{figures/armDPwithoutPotentialWorkspaceCoarse.eps}\qquad\epsffile{figures/armDPwithoutPotentialCoarse.eps}}
Problem: may be no path in pure freespace cells\\
Solution: recursive decomposition of mixed (free+obstacle) cells
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Skeletonization: Voronoi diagram}
Voronoi diagram: locus of points equidistant from obstacles
\vspace*{0.2in}
\epsfxsize=0.44\textwidth
\fig{figures/armVoronoi.eps}
Problem: doesn't scale well to higher dimensions
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Skeletonization: Probabilistic Roadmap}
A probabilistic roadmap is generated by generating random points in
C-space and keeping those in freespace; create graph by joining
pairs by straight lines
\vspace*{0.2in}
\epsfxsize=0.44\textwidth
\fig{figures/armRoadmap.eps}
Problem: need to generate enough points to ensure that every
start/goal pair is connected through the graph
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Motor control}
Can view the motor control problem as a search problem\\
in the \note{dynamic} rather than \note{kinematic} state space:\al
-- state space defined by $x_1,x_2,\ldots,\dot{x_1},\dot{x_2},\ldots$\al
-- continuous, high-dimensional (Sarcos humanoid: 162 dimensions)
Deterministic control: many problems are exactly solvable\\
esp.~if linear, low-dimensional, exactly known, observable
Simple \defn{regulatory control} laws are effective for specified motions
Stochastic \defn{optimal control}: very few problems exactly solvable\al
$\implies$ approximate/adaptive methods
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Biological motor control}
Motor control systems are characterized by massive redundancy
Infinitely many trajectories achieve any given task
E.g., 3-link arm moving in plane throwing at a target\nl
simple 12-parameter controller, one degree of freedom at target\nl
11-dimensional continuous space of optimal controllers
Idea: if the arm is noisy, only ``one'' optimal policy
minimizes error at target
I.e., noise-tolerance might explain actual motor behaviour
Harris \& Wolpert ({\em Nature}, 1998): signal-dependent noise\\
explains eye saccade velocity profile perfectly
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Setup}
Suppose a controller has ``intended'' control parameters $\theta_0$\\
which are corrupted by noise, giving $\theta$ drawn from $P_{\theta_0}$
Output (e.g., distance from target) $y = F(\theta)$;
\epsfxsize=0.7\maxfigwidth
\fig{figures/arm-setup.ps}
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Simple learning algorithm: Stochastic gradient}
Minimize $E_{\theta}[y^2]$ by gradient descent:
\begin{eqnarray*}
\nabla_{\theta_0} E_{\theta}[y^2]
&=& \nabla_{\theta_0} \int P_{\theta_0}(\theta)F(\theta)^2 d\theta \\
&=& \int \frac{\nabla_{\theta_0} P_{\theta_0}(\theta)}{P_{\theta_0}(\theta)}
F(\theta)^2 P_{\theta_0}(\theta)d\theta \\
&=& E_{\theta}[\frac{\nabla_{\theta_0} P_{\theta_0}(\theta)}
{P_{\theta_0}(\theta)}
y^2]
\end{eqnarray*}
Given samples $(\theta_j,y_j)$, $j=1,\ldots,N$, we have
\[
\hat{\nabla_{\theta_0} E_{\theta}[y^2]} =
\frac{1}{N}\sum_{j=1}^N
\frac{\nabla_{\theta_0} P_{\theta_0}(\theta_j)}
{P_{\theta_0}(\theta_j)}
y_j^2
\]
For Gaussian noise with covariance $\Sigma$, i.e.,
$P_{\theta_0}(\theta) = N(\theta_0,\Sigma)$, we obtain
\[
\hat{\nabla_{\theta_0} E_{\theta}[y^2]} =
\frac{1}{N}\sum_{j=1}^N \Sigma^{-1}(\theta_j - \theta_0) y_j^2
\]
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{What the algorithm is doing}
\vspace*{0.3in}
\epsfxsize=0.7\maxfigwidth
\fig{figures/control-noise-distort.ps}
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Results for 2--D controller}
\vspace*{0.3in}
\epsfxsize=0.75\textwidth
\centerline{\epsffile{figures/progress3-theta.ps}}
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Results for 2--D controller}
\vspace*{0.3in}
\epsfxsize=0.75\textwidth
\centerline{\epsffile{figures/progress3-theta-zoom.ps}}
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Results for 2--D controller}
\vspace*{0.3in}
\epsfxsize=0.75\textwidth
\centerline{\epsffile{figures/progress3-score.ps}}
%%%%%%%%%%%% Slide %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\heading{Summary}
The rubber hits the road
Mobile robots and manipulators
Degrees of freedom to define robot configuration
Localization and mapping as probabilistic inference problems\al
(require good sensor and motion models)
Motion planning in configuration space\al
requires some method for finitization
\end{huge}
\end{document}