ExperienceBased SurfaceDiscernment by a Quadruped Robot

1
Experience-BasedSurface-Discernment by
aQuadruped Robot

• by
• Lars Holmstrom, Drew Toland, and George Lendaris
• Portland State University, Portland, OR

2
Motivation for the Research

• Enhance our Understanding of Intelligence
• Study aspects of how our minds work by
  implementing intelligent behavior in software and
  hardware

• Application
• Create intelligent tools and algorithms to solve
  complex problems in a more human-like way

3
Desirable Human-Like Abilities

• 1. Efficient transfer of knowledge from one
  problem domain to another

4
Desirable Human-Like Abilities

• 2. Rapid Context Discernment (System
  Identification)

5
Desirable Human-Like Abilities

• 3. The more knowledge one obtains, the more
  efficient one becomes at accessing and using that
  knowledge

O(log n) search for binary trees
6
Experienced-Based Control

• Goal to build into machines the ability to use
  past experience when
• performing system identification, and
• coming up with a good controller for a given
  situation
• To do so effectively and efficiently
• To do so in a human-like fashion

7
Sony AIBO
8
Available Data

• Vision
• IR sensors
• Accelerometers
• Joint positions

9
AIBO Experience-Based Algorithm

• Goal AIBO to change gait based on surface type
• Identify change in surface
• Use only information available from joint
  actuators
• Implement proper change to gait parameters
• Recall gaits from previously-experienced surfaces
• Desire to generalize to novel surfaces
• Both tasks are to be based on past experiences
  with similar surfaces

10
How Do We Endow AIBO with Experience?

• Base capability for walking behavior is provided
  by Sony (hardware) and Tekkotsu (software)
• Train AIBO to develop good gaits (gait
  parameters) for a selected set of distinct
  surfaces

11
Genetic Algorithm Used to Develop/Learn Gaits

• Optimized for a balance of speed and sway on 4
  different surfaces
• 1. Hardboard
• 2. Thin foam
• 3. Thin carpet
• 4. Shag carpet
• Each of these gaits performed significantly
  better than the default Tekkotsu gait

(Chromosome)
12
Context Discernment

• Our system now has Experience
• Good GA gaits for a set of surfaces.
• Now, how do we get AIBO to recognize and then
  adapt to changes in surface qualities?

13
Available Data

• Vision
• IR sensors
• Accelerometers
• Joint positions

14
Observation Issued Commands and Measured Motion
Differ
15
Are There Measurable Differences Between the
Kinesthetic Responses for Different Surface Types?
16
Complicating Attributes of Available Data for
this Task

• Low sampling rate (31Hz)
• Occasional dropped samples
• Large variance
• Process noise?
• Measurement noise?
• Time consuming to collect
• Non-stationarities
• Surface Irregularities
• Physical Dynamics of the AIBO

17
Approach 1 Work in the Frequency Domain

• Smoothed Periodograms of Left Hip Joint,
  Thin-Foam Gait, on 4 Different Surfaces

18
Approach 2 Work in the Time Domain

• Linear Forward Prediction Model
• For each of the 15 actuator signals, predict the
  current state of the actuator as a linear sum of
  the actuators past states.
• Find the mean squared error (MSE) of the
  predictions across all of the actuators at each
  time step.
• Using this procedure for a single gait/surface
  combination and set of actuator signals, we can
  generate a one-dimensional error signal

19
Fitting the Models

• Fit one model on the data collected for each
  gait/surface combination
• Solution of the Normal Equations is performed to
  quickly find the unique and optimal model
  parameters for the given training data
• Above properties motivated use of linear models
  (for computational ease)

20
Kinesthetic Experience(over all modeled
gait/surface combinations)

• The Kinesthetic Experience is the (gaits
  surfaces) dimensional signal indicating each
  models MSE as it unfolds over time
• In figure, error signal corres-ponding to the
  actual gait being used and surface being
  traversed is the minimum in each of these cases,
  indicating perfect classification

21
Implemented Discernment of Surface Transition

• The algorithm can discern surface transitions in
  novel data within 2-4 seconds with 92 accuracy
• Accuracy increased as time increases

22
The Same Approach Can Be Applied to Discerning
Changes in Surface Incline
Surface Transition Discernment
Surface Discernment
23
Experience Based Discernment and Control In Action
24
Future Directions

• Use a set of gait/surface experiences as the base
  for discerning novel surface types
• The Kinesthetic Experience acts as a parametric
  description of the surface being experienced
• Use the Kinesthetic Experience to generalize to
  new control policies to match new surfaces