Nicolas Galoppo von Borries

1
Neural Navigation Approach for Intelligent
Autonomous Vehicles (IAV) in Partially Structured
Environments

• Chohra and C. Benmehrez
• A. Farah

2
Neural networks for navigation

• Neural networks advantages
• Robustness towards noisy data, thus well suited
  for sensorial data processing
• Not rule-based handle wide palette of
  environments
• Unknown
• Complex
• Dynamical

3
Neural networks robustness

• Short lifespan of sensors
• Knowledge in NN is distributed over several
  neurons, therefore there is redundancy
• Noisy data
• Classical algorithms fail
• too many cases to handle

4
Motion planning environment model

• Dynamic
• Learning from examples
• Online learning continuous weight updates
• Unknown / unexplored
• Learning from examples, able to generalize (this
  is interpolation/extrapolation)
• Complex
• No large programming effort and cost
• Domain expertise not needed, use training data

5
Even more advantages

• Neural networks are inherently parallel
• Parallel implementation is easy
• Real-time response adds to robustness because
  late reactions lead to corrective measures

6
Compared to classic approaches

• Roadmaps and cell decomposition
• Construct data structure to model configuration
  space large data structures and complex
  geometric computations
• Advantage can be reused for multiple-query
• Potential field methods
• In a way, similar to neural network approach
• Expensive preprocessing and difficult update
  inefficient for dynamic environments

7
Motion planning with Artificial Intelligence
techniques?

• Neural navigation systems are based on the way
  humans react to their environment

8
Application of Motion Planning

• Motion of an intelligent autonomous vehicle in a
  unknown, dynamic environment
• Known
• Angle towards target
• Obstacles in local environment
• Global environment
• Dynamic and unknown
• Data
• Low-quality sensors, noisy, occasional failures
• Obstacles
• Static, Intelligent (other vehicles),
  non-intelligent

9
System overview

• The environment is complex, the system design
  must reduce complexity
• How?
• Reactive look at local environment only
• Hierarchical structure

10
The system hardware

• Vehicles 5 directions of movement
• 5 possible actions Ai go forward in direction i
  (angle 30, 60, 90, 120, 150)
• Sensors 5 ultrasonic sensors
• detect the local vehicle-obstacle configuration
• range of 0.15-10.5 m
• 15 degree angle beam

11
Hierarchical structure of the NN

• Three subtasks
• NN1 target localization T
• NN2 obstacle avoidance O
• NN3 coordinator decides action A

12
Target localization NN1

• Its a classifier, the output is a vector
  indicating confidence values for the sectors T1
  T6
• Input vector XT (X1,,X5), defines a temperature
  field, to indicate the direction of the target
• NN1 must learn which sector to choose, given
  input temperature vector XT

13
Obstacle avoidance NN2

• Also a classifier, the output is a particular
  local obstacle configuration (O1 O30)
• Input vector XO(X1,,X5), Xi denotes minimum
  distance to an obstacle at angle i

14
Decision-making NN3

• Its the coordinator
• Combines outputs of NN1 and NN2
• 5 output neurons, the firing one indicates the
  action Ai for the vehicle

15
Training

• We can train NN1, NN2 and NN3 separately, because
  of network hierarchy
• Speeds up learning
• Less training examples needed
• Back propagation of the error at the output
• Performed incrementally, one example vector at a
  time, the network learns how to adapt to unknown
  situations
• Well suited for non-stationary environments

16
Training of NN1

• Obstacle free simulation environment
• The vehicle moves along different paths around
  the target, traversing different vehicle-target
  orientations and positions
• For each vehicle target configuration, a
  desired position class Ti is assigned for error
  back propagation

17
Training of NN2

• Given a fixed vehicle target configuration,
  various obstacles are placed in the environment
• Each training sample is assigned one of the 30
  possible obstacle configuration classes

18
Backpropagation in NN1 and NN2

• Initialize weights
• Apply input vector X
• Compute outputs Y of hidden layer
• Compute outputs C of output layer
• Compute error
• Update weights
• Repeat until errorltthreshold

19
Training of NN3

• Reinforcement learning by trial-and-error
• Penalty reward system is used
• Target localization penalty a human critic
  penalizes target localization
• Obstacle avoidance penalty sensor data
  determines the distance to the closest object in
  each direction
• Plocalization ltlt Pobstacle avoidance, collision
  must be avoided at all cost

20
Simulation results

• Neural network approach provides robustness
• Real time performance, for simple as well as
  complex environments
• Static, intelligent and non-intelligent moving
  obstacles are successfully avoided

21
Static environments
22
Intelligent obstacles
23
Non-intelligent obstacles
24
Complex environments
25
Conclusions

• Neural network approach to motion planning
• Advantages
• Powerful adaptive learning real-time performance
  in unknown environments
• Ease of use little domain-specific expertise
  needed
• Built-in robustness due to redundancy, the
  network capabilities may be retained with network
  damage
• Disadvantages
• Dead-end situations are hard to avoid

26
Hybrid methods

• Memory elements allow for long-term behavioral
  patterns
• E.g. avoidance by reversing path or stopping
  briefly
• Output a set of actions helps avoiding dead-end
  situations (i.e. local minima)
• Use of genetic algorithms deduce the optimal
  network structure
• Improves learning performance
• Disadvantage requires domain specific expertise,
  which is not needed for hierarchical structure