SENSOR BASED CONTROL OF AUTONOMOUS ROBOTS

1
SENSOR BASED CONTROL OF AUTONOMOUS ROBOTS

• Robert Mahony
• Department of Engineering
• Australian National University

Department of Engineering Research Forum,
November, 2006.
2
Active topics in robotics

• Physical control of vehicle/machine
• Sensory perception of environment
• Interaction of vehicle with environment
• AI (artificial intelligence)

Autonomy
Co-ordination
Data interpretation

• Multiple robots
• Multiple tasks
• Human-robot interfaces
• Inter-robot interfaces

• Data fusion
• SLAM (simultaneous localisation and mapping)
• Object recognition, sensor segmentation

3
Active topics in robotics

• Physical control of vehicle/machine
• Sensory perception of environment
• Interaction of vehicle with environment
• AI (artificial intelligence)

Autonomy
SENSOR BASED CONTROL
Co-ordination
Data interpretation

• Multiple robots
• Multiple tasks
• Human-robot interfaces
• Inter-robot interfaces

• Data fusion
• SLAM (simultaneous localisation and mapping)
• Object recognition, sensor segmentation

4
Autonomy in robotic systems

• An autonomous robot is capable of moving about
  within an unstructured (or partially) structured
  environment independently.
• Unstructured environment
• No map available.
• Partially structured environment
• There is a map but it does not contain all
  objects and is not necessarily accurate.
• In all cases the robot must regulate its motion
  with respect to the local environment.

5
Example Aerial robot
A common task that is considered in aerial
robotics is regulation of the vehicle relative to
an observed feature.

• Other important tasks
• Obstacle avoidance
• Close approach and landing

6
Classical control approach

• Classical control theory provides a standard
  approach to regulation problems
• Model the dynamics of the system.
• Represent the dynamics in terms of a minimal
  state.
• Represent the task in terms of a state error.
• Design a control algorithm to drive the state
  error to zero.
• Measure something.
• Estimate the system state on-line.
• Input the state estimate into the control
  algorithm to close the loop.

7
Issues with classical control approach
REAL WORLD
Observations
Sensors

• The mapping from observation to state estimate is
  
• non-linear
• over-determined
• ill-conditioned

State estimates
Task error

• Computing a state estimate from the observations
  requires
• Model of the environment (SLAM)
• Model of the system dynamics
• Estimates tends to be ill-conditioned when the
  vehicle is distant from local features.

Task error is naturally conditioned relative to
proximity to environment! Easy to represent in
terms of sensor measurements.
8
Sensor based control

• Sensor based control is a paradigm that is only
  subtly different from the classical approach.
• Model the dynamics of the system
• Use this model to determine the dynamic response
  of the sensor signals based on the expected
  environment.
• Represent the task in terms of a sensor error
• Design a control algorithm to drive the sensor
  error to zero based on analysis of the sensor
  dynamics
• Input the sensor measurements into the control
  algorithm to close the loop.

9
Bio-mimetic systems

• One of the major motivations for sensor based
  control of autonomous robots is the growing
  evidence for simple sensor based control
  algorithms in biological system.

A honey bee regulates its thrust in landing
approach in proportion to a measure of
divergence of the observed optic flow (Srinivasan
et al. 2000, Moffit et al. 2006)
Optical flow field ? of textured surface under
direct approach
10
Challenges to sensor based control.

• Sensor dynamics tend to be highly non-linear.
• Very challenging control problems.
• Sensor data tends to be high dimensional much
  higher dimensional than the state vector.
• Leads to non-minimal system representations.
• Early work in this area has depended on finding
  good features (eg average flow divergence s? div
  ?) that provide a low dimensional sensor state
  representation.
• Overcoming these problems leads to highly robust
  and effective task based control of autonomous
  systems.

11
Stabilisation of aerial robot relative to image.
Observed closed-loop error evolution in the
sensor based task criterion.
Regulation of position in task space. Computed
from inverse pose algorithm.
12
Collaborators
Peter Corke
Tarek Hamel
Odile Bourquardez
Nicolas Guenard
(many other honours and stagiere students)
Francois Chaumette
13
Dynamic image based visual servo control

• Consider the problem of stabilising an aerial
  robot relative to some physical object whos
  image is easily segmented.

Observed object
Image on spherical image plane
Spherical centroid is the integral of observed
image on the sphere.
14
Sensor space dynamics and control