Below is a list of some project ideas for the class. You're also welcome
to propose your own STAIR-related project. We'll pick projects in the
class' first meeting.
Perception (vision) for robotic manipulation.
In robotics, it is a fairly "easy" problem to pick up many objects so long as
we have a detailed 3D CAD model of the object. However, if we are seeing a new
object for the first time through our camera, it is then a very hard problem
just to figure out how to pick the object up. In this project, we will develop
computer vision algorithms for STAIR that learn to to pick up any ordinary, home
or office object (such as staplers, cups, pens, a hammer, pliers, etc).
Spoken dialog for STAIR.
We want STAIR to be able to understand and execute verbal/spoken commands such as:
"Please invite Daphne to join us for the meeting."
"Would you please see if Dan is in his office?"
"Please tell Sebastian we'll be 5 minutes late."
"Let Ken know that we're done with the report."
In this project, we'll develop a spoken dialog system that can understand
such commands. The emphasis of this project will be the speech understanding
(rather than speech recognition).
Indoor robot navigation.
We want STAIR to be able to safely go anywhere within the Stanford Gates building.
For this, one of the most crucial capabilities for the robot will be efficient and
precise navigation in close proximity to obstacles. Compared to most indoor robots, we
will require significantly more precise navigation and localization, to enable tasks
such as manipulating (opening) a door handle. In this project, we'll implement high-accuracy
localization and navigation algorithms for STAIR.
Mapping the Gates building.
In order to go anywhere within the Gates building, STAIR will need a map of the building
that is annotated with objects (such as doors and elevators that it can use). This project
will focus on mapping out (using STAIR's laser scanner) the Gates building, creating precise
models of objects, and annotating the Gates map with these objects.
Opening doors and pushing elevator buttons.
In order to go anywhere in Gates, STAIR will need to be able to open doors and
push elevator buttons. We will develop robotic manipulation algorithms for opening
doors and pushing elevator buttons. In either case, the basic approach will be to use
either computer vision or the laser scanner to localize the robot precisely relative to a
specific door handle or the elevator button, and then reaching out to perform the manipulation
task.
Computer vision: Fast, robust object recognition.
We would like STAIR to understand a large vocabulary of objects in everyday home
and office environments. To enable this, we will use a "cascade" (linear chain) of classifiers
that quickly rejects pictures that do not contain any interesting object. We will implement
the Wu/Rehg cascade of classifiers, and extend it to multi-class object recognition
for STAIR.
Computer vision: Segmentation from parallax.
Segmentation is often considered a key step in object recognition and scene understanding. We will
develop a novel algorithm that uses stereo (two cameras) and/or parallax (different images resulting
from motion of the robot) to create a basis for segmentation. Unlike traditional segmentation
algorithms, we will therefore exploit cues other than color to perform the task.
Seeing itself.
For an intelligent robot, part of understanding the perceived images means being able to
see and understand images of itself---specifically, its own hand/arm. We will develop a
reliable system for recognizing the robot's own arm in the video. This will be done
using both kinesthetic information about the position of the arm, as well as the
appearance of the arm as it moves around in the image.
Propose your own.
You are also very welcome to propose any other research project that is relevant to STAIR.
