MICHAEL MILFORD, DAVID PRASSER, AND GORDON WYETH FOLAMI ALAMUDUN GRADUATE STUDENT COMPUTER SCIENCE & ENGINEERING TEXAS A&M UNIVERSITY RatSLAM on the Edge:

MICHAEL MILFORD, DAVID PRASSER, AND GORDON WYETH

FOLAMI ALAMUDUN

GRADUATE STUDENTCOMPUTER SCIENCE & ENGINEERING

TEXAS A&M UNIVERSITY

RatSLAM on the Edge: Revealing a Coherent Representation from

an Overloaded Rat Brain

OUTLINE

OverviewRatSLAMExperience MappingGoal Recall Using Experience MapsExperiment ResultsDiscussion

OVERVIEW

In order for a robot to navigate intelligently :

It must possess a means of acquiring and storing information about past experiences; and

It must possess the ability to use make decisions based on this information.

OVERVIEW

What is SLAM?Simultaneous Localization and

Mapping

Determine the state of the world:

What does the world look like?

Determine location in the observed world:

Where in the world am I?Where in the world…?

OUTLINE

OverviewRatSLAMExperience MappingGoal Recall Using Experience MapsExperiment ResultsDiscussion

ratSLAM

Why are we SLAM-ing?Maps are used to depict the environment for

an overview and to determine location within the perceived environment.

Locating and mapping under conditions of errors and of noise is very complex.

Simultaneous localization and mapping (SLAM).

Inspired by computational models of hippocampus in rodents.

Hippocampus is a part of the brain that plays an important role in long-term memory and spatial navigation

Neurons in the rat and mouse hippocampus respond as place cells.

Place cells exhibit a high rate of firing whenever an animal is in a location in an environment corresponding to the cell's "place field"

Place Field are patterns of neural activity that correspond to locations in space

ratSLAM

ratSLAM

RatSLAM is an implementation of a hippocampal model of robot control:

To provide a new and effective method for the mobile robot problem of (SLAM); and

To reproduce a high-level brain function in a robot in order to increase the understanding of memory and learning in mammals.

ratSLAM

Architecture for RatSLAM.

Local View and Pose Cell arrangement for artificial landmarks

ratSLAM – Local View

Local View (LV):A form of representation processed from vision

information from camera imagesCalibrates the robot’s state informationStored and associated with the currently

active pose cells.If familiar, the current visual scene also

causes activity to be injected into the pose cells associated with the currently active LV cells.

ratSLAM – Pose Cell

3-D pose cell model. Each dimension corresponds to one of the three state variables of a ground-based robot

Θ`

y`

x`

ratSLAM – Pose Cell

Pose Cell:A three-dimensional structured competitive

attractor neural network.Combines the characteristics of place and

head-direction cellsEach axis of the structure corresponds to a

different state variable, x′, y′ and θ′

ratSLAM

How it works:Wheel encoder information is used to perform

path integration by shifting the current pose cell activity.

Vision information is converted into a local view.

Local view cell is associated with the currently active pose cells.

If familiar, activity is injected into the particular pose cells associated with the currently active local view cells.

ratSLAM – Pose Cell

The first test environment was a two by two metre arena

ratSLAM – Pose Cell

Floor plan and robot trajectory for initial goal navigation experiments.

ratSLAM – Pose Cell

The temporal map cells after recall of the first goal.

ratSLAM – Pose Cell

The path the robot followed to reach the first goal.

ratSLAM

Hashing collisions in the Pose Cells:Vision information starts to cause more

frequent loop closures. Leads to discontinuities in the pose cell

matrix.Multiple representations of the same physical

areas in the environment.Clusters of pose cells become associated with

more than one pose.Hashing collisions within pose cells are

unavoidable.

ratSLAM

Floor plan of large indoor environment

ratSLAM

Dominant packet path for a 40 × 20 × 36 pose cell matrix.

ratSLAM

Temporal map for the large indoor environment

OUTLINE

OverviewRatSLAMExperience MappingGoal Recall Using Experience MapsExperiment ResultsDiscussion

EXPERIENCE MAPPING

Experience mapping algorithm is the creation and maintenance of a collection of experiences and inter-experience links. 

This produces a spatially continuous map without collisions from the messy representations found in the pose cells.

It does this by combining information from the pose cells with the Local View cells and the robot's current behavior

EXPERIENCE MAPPING

Experience map co-ordinate space

EXPERIENCE MAPPING

Experience Mapping:The algorithm uses output from pose cells

and local view cells to create an experience map.

A graph-like map containing nodes (experiences) and links between experiences.

Each node represents a snapshot of the activity within pose cells and local view cells.

New experience nodes is created as needed.

EXPERIENCE MAPPING

Experience GenerationEach experience has its own (x′, y′, θ′, V ).

where x’, y’, and ′ are the three state variables. V describes the visual scene associated with the experience.

Output from the pose cells and local view cells is used to create a map made up of robot experiences.

Inter-experience links store temporal, behavioral, and odometric information about the robot's movement between experiences.

EXPERIENCE MAPPING

Experience zone of influence:Activity is dependent on how close the activity

peaks in the pose cells and local view cells are to the cells associated with the experience.

EXPERIENCE MAPPING

x'PC, ,YPC and θ' - coordinates of the dominant activity packet, x'i, yI, and θ‘ - coordinates of the associated experience i, ra is the zone constant for the (x',y') plane, and0a is the zone constant for the 0' dimension.

EXPERIENCE MAPPING

Experience zone Visual scene

V is the current visual scene. Vi is the visual scene associate with

experience i.Ex’y’θ’ is the visual scene energy component.

EXPERIENCE MAPPING

Total Energy Level:

Total Energy level of Experience Ei:Ei = EV × (Exy + Eθ)

EXPERIENCE MAPPING

Experience Mapping:As the robot moves around a novel

environment, it needs to generate experiences to form a representation of the world.

Learning of new experiences is triggered not only by exploring in new areas of an environment, but also by visual changes in areas the robot has already explored.

OUTLINE

OverviewRatSLAMExperience MappingGoal Recall Using Experience MapsExperiment ResultsDiscussion

GOAL RECOLLECTION

Experience Transitions:

Transitions represent the physical movement of the robot in theworld as it moves from one experience to another.

GOAL RECOLLECTION

Experience Transitions:

dpij is a vector describing the position and orientation of experience j relative to experience i.

GOAL RECOLLECTION

Map Correction:Discrepancies between a transition's odometric informationand the linked experiences' coordinates are minimized through a process of map correction:

OUTLINE

OverviewRatSLAMExperience MappingGoal Recall Using Experience MapsExperiment ResultsDiscussion

EXPERIMENTAL RESULTS

EXPERIMENTAL RESULTS

EXPERIMENTAL RESULTS

EXPERIMENTAL RESULTS

DISCUSSION

Experience maps are localized: Cartesian properties are not guaranteed beyond

local areaFor instance straight corridors may be

slightly curved in the experience map.