NeoN: Neuromorphic Navigation with DANNA · Example 15x15 DANNA Network LIDAR 2 LIDAR 3 LIDAR 4 LIDAR 5 SWITCH 1 SWITCH 2 Left Motor FW Left Motor BW Right Motor BW Right Motor FW

NeoN: Neuromorphic Navigation with DANNA

J. Parker Mitchell, Grant Bruer, Mark E. Dean

Overview

● EECS Senior Design Project● Create roaming robot

○ Avoid obstacles○ Avoid ledges

● Design a physical framework for deploying DANNA

● Explore robotics control as a problem domain

Robot Design

● Tank style drivetrain● LIDAR mounted on servo● Limit switches● All computation through FPGA

30°30° 30°

30°

Network I/O

Lidar Reading (meters)

Fire Weight

LIDAR Fire Weight Conversion

● 9 Inputs○ 5 LIDAR readings○ 2 switches○ Bias○ Random

● 4 Outputs○ Left motor forward/backward○ Right motor forward/backward

Network I/ONeuronSynapseOutputInput

Legend

Example 15x15 DANNA Network

LIDAR 2

LIDAR 3

LIDAR 4LIDAR 5

SWITCH 1

SWITCH 2

Left Motor FWLeft Motor BW

Right Motor BWRight Motor FW

RANDOM

BIAS

LIDAR 1

Training● Train using evolutionary optimization (EO)● Simulate in an empty room, room with

obstacles, and table with obstacles● Score based on grid coverage● Penalize for critical failures

○ Hitting an obstacle○ Falling off ledge

Grid Coverage Example

Simulation Video

Future Work

● DANNA with leaky IAF neurons● Nonsquare networks, input spacing● Optimized FPGA logic● Harder task

○ Target tracking○ Autonomous flying drone

● Implementation with mrDANNA

Any Questions?

Visit our website at neuromorphic.eecs.utk.edu