Intentional Motion Online Learning and Prediction

1
Intentional Motion On-line Learning and Prediction

• December 2005
• Dizan Vasquez, Thierry Fraichard, Olivier Aycard
  and Christian Laugier
• e-Motion, GRAVIR, CNRS
• INPGrenoble, INRIA, France
• firstname.lastname_at_inrialpes.fr

2
Objective

• To predict intentional motion on the basis of
  past and present observations.
• How?Observe the environment in order to learn
  typical behaviors, and then use them to predict.

3
Related Works

• Trajectory Prototypes.KruseEtAl97,BennewitzEtAl0
  2,VasquezFraichard04
• Learning using clustering
• Prediction using similaritymeasure.
• State Based Models.Zhu91,BuiEtAl02
• Learning using EM.
• Prediction using bayesianinference.
• Problems
• Unable to learn newbehaviors.
• Unable to adapt to achanging environment.
• Slow.

4
Proposed Approach Plan Representation

• Learn and Predict HMM STRUCTURE PARAMETERS
  GOAL IDENTIFICATION PREDICTION
• An HMM is used to represent a planp( qt, qt-1,
  ot ) p( qt-1 ) p( qt qt-1 ) p( ot qt )

qt-1
qt
qt1
ot-1
ot
ot1

• Extending to many plans p( qt, qt-1, ot, ? )
  p( qt-1 ) p( ? ) p( qt qt-1, ?,) p( ot qt )

?
qt-1
qt
qt1
ot-1
ot
ot1
5
Proposed Approach Learning

• Parameters to be learned
• Number of states N.
• Number of goals G.
• Allowed Transitions.
• Transition Probabilities p( qt qt-1, ?,).
  Table Ag,i,j.
• Observation Probabilities p( ot qt
  ).Gaussians Gi(µi,si)
• Learning Process

Structure
Parameters
Prediction
Prediction
Prediction
O x1, y1 , x2, y2 , , xT, yT
Transition Learning
State Learning
State Learning
State Learning
Goal Learning
6
Proposed Approach Learning (continues)

• State learning withGrowing Neural Gas
  Fritzke95
• Place states where things happen.
• Advantages
• Fast (O(N) in time).
• N not fixed a priori.
• Incremental.
• Builds a planar graph

• Gaussian parameters
• µi wi.
• si average neighbors distance / 2.

7
Proposed Approach Learning (continues)

• Goal Learning
• Another GNG is used.
• Only terminal observations are taken into
  account.
• Transition probabilities
• g Nearest goal to trajectory end.
• Counts are updated on a maximum
  likelihood/Viterbi criterion.

8
Proposed Approach Prediction

• Exact inference is possible in real time!
• Belief-state update O(GN2)? O(GN).
• Prediction modes
• Final goal. O(N)
• Intermediate states. O(GN2) ? O(GN).

9
Experimental Results (continues)
10
Conclusions and further work

• Our intentional motion approach
• Represent motion plans as HMM.
• Is able to learn and predict simultaneously.
• Proposes a novel use of GNG to determine the
  number of HMMs, as well as their structure and
  parameters.
• Thanks to the chosen learning approach, exact
  inference is linear with respect to the number of
  states.

11
Further Work

• Work has started on the parking lot
  environment.
• Problems detected with GNG for this kind of
  environment. Switched to Grow When Required
  MarslandEtAl02.

12

• Thank You!

13
Experimental Results
14
Proposed Approach Prediction (continues)
15
Objectives

• To predict intentional motion (ie motion
  resulting from plan execution) on the basis of
  past and present observations

?
?