Week 13 Overheads

Modelling Navigation in Computer or Robotic Agents

---

Rodney Brooks and Company at MIT

Hannibal

A robot patterned very closely on insects.

 

Pebbles

Pebbles navigating in an outdoor environment.

 

Pebbles avoiding a downed...obstacle.

 

Ants

Front and back view of an ant robot. Note that thexe robots are about 3 cm long (take this link for a better idea of the scale).

 

A lot of sensors, emitters, and detectors are packed onto an ant robot.

 

     

An example of behaviour in a group of ants. In the figure on the left, the centre ant has found a "food item". When this ant picks up the food in its mandibles it emits an "I've found food!" signal that the other ants in the surrounding area can detect. In the figure on the right the other ants move in to see if there is any more food available.

---

Yamauchi and Beer

ELDEN system

Evidence grids
Correct for dead reckoning errors

     

---

Topological Mapping

Grids and Voronoi diagrams

     

Regions and topological graph

     

Pruned regions and topological graph

     

---

Wittmann & Schwegler (1995)

Are shortcuts necessary for ant path integration?

Simple, logical assumptions

Compass orientation

• Circular array of dedicated neurons

Distance

• Firing rate of neuron "clock"

Short-term memory store

Accurage path integration

Alternate solution?

• Catch features

---

Waterstrider Prey Orientation Model: Snyder (1998)

Simple artificial neural network

---Insert Image---

Baseline

Amputations

Relearning?

Performance of the ANN when the receptors of the first and second right legs are lesioned (left) and after re-training (right). Notice that prior to re-training the ANN makes the expected systematic errors in rotation (i.e., rotations to the left instead of the right within the angular range covered by the lesioned receptors). Following re-training the lesioned network performs as would be expected if it was intact.