- inspired from rat (or rodent) brain hippocampus
- leads to large scale navigation
- brain cells used for navigation spotted in rat brains:
- place cells (cells that are activated only when the rate is at a given place)
- head direction cells (fires when the head is oriented in a specific direction)
- the brain maintains a code that describes a 3dof pose that is maintains through visual and proprioceptive inputs
Heading Attractor (head direction cells)
- Continuous Attractor Network to model brain cells: with spreading connections and inhibition connections
- Self Motion cues add connections in the network
- Local views cells that are associated to a specific orientation and fires to it when activated
2D attractor (place cells)
- boundary problem solved with wrapping connectivity
First Exp (2003)
- 2m * 2m square with artificial landmark
- the error keeps low in the short term and diverge in the long run: because of ambiguity of pose cells for some observation that must lead to multi hypothesis tracking
Version 2
- now use a combination of place cells and orientation cells as pose cells
- But in 2005, new conclusions in biology: a single place cell can be activated on several places as well as conjuctive cells => pose cells organised in a grid
- experiments in a building: lots of relocalization that leads to discontinuities => the network cannot be interpreted as locations anymore
Version 3
- add an experience map = semi metric map
- desambiguate distinct scenes that looks similar
- experience map = graph : graph relaxation for loop closure, pruning to limit memory, path planning
Suburb experiments
- a raw VO gives the motion info
- 42 loop closure detected
- not metrically correct, but the topology is well represented
Conclusion
- start with biology, then engeneer where it fails