Transcript jpineau-CogSciResearchDay-sep10

Cognitive Robotics: Lessons from the
SmartWheeler project
Joelle Pineau, [email protected]
School of Computer Science, McGill University
September 22, 2010
Cognitive robotics
• Main scientific goal: Design robots that exhibit intelligent
behavior by providing them with the ability to learn and reason.
Abilities
Goals/Preferences
Prior Knowledge
Robot
Observations
Actions
Environment
• Main tools: Probability theory, statistics, optimization, analysis
of algorithms, numerical approximations, robotics, …
2
Joelle Pineau
Why build the SmartWheeler?
• Potential to increase the mobility and freedom of individuals with
serious chronic mobility impairments is immense.
– ~4.3 million users of powered wheelchairs in the US (Simpson, 2008).
– Up to 40% of patients find daily steering and
maneuvering tasks to be difficult or impossible
(Fehr, 2000).
• An intelligent wheelchair platform provides
opportunities to investigate a wide
spectrum of cognitive robotics problems.
3
Joelle Pineau
The robot platform
1st generation (McGill)
•
Standard commercial wheelchair.
•
Onboard computer and custom-made electronics.
•
Sensors: laser range-finders, wheel odometers.
•
Communication: 2-way voice, touch-sensitive LCD.
2nd generation (Polytechnique)
4
Joelle Pineau
Software architecture
Two primary components of cognitive robotic system:
Interaction Manager and Navigation Manager
5
Joelle Pineau
Reinforcement learning paradigm
T 
Choose actions such as maximize the sum of rewards, E  rt 
t 0 
6
Joelle Pineau
Navigation management
7
Joelle Pineau
Autonomous navigation software
8
Joelle Pineau
Variable resolution robot path planning
9
Joelle Pineau
Interaction Management
10
Joelle Pineau
User interaction example
11
Joelle Pineau
Wheelchair Skills Test
http://www.wheelchairskillsprogram.ca
• Set of 39 wheelchair skills developed to test/train wheelchair users.
– Each task graded for Performance and Safety on Pass/Fail scale.
• Allows comparison and aggregation of results.
12
Joelle Pineau
Voice interaction with healthy subjects
13
Joelle Pineau
Voice interaction with target population
14
Joelle Pineau
Qualitative analysis
• Positive
– Impressed by autonomous functionality
– Obstacle avoidance
– Visual feedback
• Negative
– Wanted more time to familiarize with the system
– Too much micromanagement
– Microphone required on/off button
15
Joelle Pineau
Discussion
• Current experimental protocol is constrained.
• Useful for formal testing, inter/intra-subject comparison.
• Limited use for measuring long-term impact.
• Extension to standard living environments is possible.
• Navigation in indoor living environments is possible.
• Navigation in outdoor or large indoor environments is challenging.
• Communication is reasonably robust for most subjects.
• But suffers from lag, noise, and other problems.
• Multi-modal interface is desirable but harder to design.
• Need to investigate life-long learning for automatically adapting
to new environments, new habits, and new activities.
16
Joelle Pineau
Project Team
• McGill University:
– Amin Atrash, Robert Kaplow, Julien Villemure, Robert West, Hiba
Yamani
• Ecole Polytechnique de Montréal:
– Paul Cohen, Sousso Kelouwani, Hai Nguyen, Patrice Boucher
• Université de Montréal
– Robert Forget, Louise Demers
• Centre de réadaptation Lucie-Bruneau
– Wormser Honoré, Claude Dufour
• Constance-Lethbridge Rehabilitation Centre
– Paula Stone, Daniel Rock, Jean-Paul Dussault
• Institut de réadaptation en déficience physique de Québec
– François Routhier
17
Joelle Pineau
Reasoning and Learning Lab, SOCS
18
Joelle Pineau
Adaptive deep-brain stimulation
Goal: To create an adaptive neuro-stimulation system that can maximally
reduce the incidence of epileptiform activity.
19
Joelle Pineau