Rover Fault Monitoring

1
Rover Fault Monitoring

• Life in the Atacama Project WorkshopJuly 28 2003
  at 1200pm
• Vandi VermaPh.D. Student, RI
• vandi_at_ri.cmu.edu

2
What is Fault/Health Monitoring?

• Simple definition
• Detecting significant deviations from expected
  behavior
• Usually passive
• Challenges
• The expected behavior in different terrain may be
  very different
• There are no sensors to directly measure robot
  failure
• The sensors are very noisy and sometimes fail
• Requires modeling complex rover dynamics

3
What can Health Monitoring do for this project?

• Keep the robot safe
• Detect situations that potentially could lead to
  faults
• Increase degree of autonomy
• Reduce operator stress
• Increase the length of hands-off traverse

4
How was health monitoring done in the first field
season?

• A set of 33 simple binary faults were considered
• Examples are
• Roll/pitch limit exceeded allowable threshold
• Missed plan waypoint missed
• Battery power low
• No arcs found by the navigator
• No heartbeat received from vehicle controller
• Min/max constraints
• Simple filtering
• Response to faults was an emergency stop unless
  overridden by operator

5
How well did the Health Monitor perform?

• Distance between fault induced stops during
  autonomous operation
• Mean 297.17 m
• Max 1167 m
• Variance 90613.97
• This ignores vehicle stops due to manual
  intervention
• Manual interventions for safety reasons 10
• Blown fuse
• High winds
• Stuck/slipping
• Manual interventions for faults we currently
  detect 3
• High pitch
• Steering axle stuck in hard left turn

6
Frequency of faults
7
Faults Detected
8
False detections
9
Lessons learned

• Current method does not handle sensor noise well
• Without autonomous recovery from faults, the
  health monitor reduces the length of hands-off
  traverses by signaling emergency stops.
• Explicit experiments to build models are required
• Our initial roll/pitch thresholds were too
  pessimistic
• Essential to detect wheel slip
• Adjustable risk

10
Implications

• Rover Design
• Better recovery
• Autonomous recovery
• More intuitive interface to faults for operator
• Probabilistic fault detection
• Higher fidelity models
• Improve accuracy
• Freeze rover design early
• Extensively test models before the field trip
• Include more information
• Low level data from controller
• Improve filtering
• Pop-up box for explanation of each Estop

11
Implications

• Future Field Investigations
• Explicit parameter estimation period in the
  beginning
• Effective method for recording false negatives
• Identifiable message source
• Pop-up box for explanation of each Estop