For Autonomous Navigation on a Small Robot
Junior Research associate 2021
Dexter Shepherd, Computer Science with Arcial Intelligence BSc
Supervisor—Dr James Knight
A problem in robocs is the high cost
of chassis design and building. The
Nanosaur combats this issue as the
chassis is fully 3D printed. This design
is open source, which allowed us to
modify it for panoramic imaging us-
ing a Rhino horn on the front with a
raspberry pi camera and reecve
dome at the summit. Mounted on
the back is Jetson Xavier board,
which runs Ubuntu.
The tracks are printed in a dierent lament to the
rest of the chassis to allow bend and exibility.
2
Reinforcement learning is a method of evolving weights and biases in a
neural network rather than training them. This is helpful when you do not
have training data and want
an agent to get very good at a
specic task. The Microbial
algorithm was used as opmi-
zaon. Fitness was deter-
mined based on the me it
took divided by the total me
given. The trial ends when the chassis
bumps into an object.
We use this method to take in visual infor-
maon from the panoramic imaging and
output motor instrucons.
Before entering the model, the image is
converted to an opcal ow representaon,
which shows the change in structured light. This helps nd where obstacle
may lie. This is achieved using the Lucas-Kanade method seen below.
Using simulaon we are able to test the model faster with more trials. Visual
input is replaced
by radar. Each
agent is given 20
seconds.
Motor control is
replaced with a
vector repre-
5
By generaon 20 nearly all trials had found the local opmal soluon. These
soluons usually found travelling in straight lines. When we change the envi-
ronment each generaon, we sll get straight lines.
The Nanosaur had many hardware issues due to the plasc easily wearing
down, parcularly round the motors where the most heat was given out.
The tracks were replaced by wheels to combat this. The algorithm in itself
worked by nding simple routes without crashing.
Going forward, having a cheap small robot would allow other researchers in
the AI group to build more and use them to invesgate swarm navigaon
where dierent robots explore dierent parts of an environment and com-
pare their informaon.
[1] Justin Wildrick Hart. Robot self-modelling.
[2] Hod Lipson Josh Bongard, Victor Zykov. Resilient machines
through continuous self-modelling.
[3] Nanosaur instructions. https://nanosaur.ai/quick-start/.
[4] Jetson nano. https://www.nvidia.com/en-gb/autonomous
-machines/embedded-systems/jetson-nano/.
[5] Melanie Schranz. Swarm robotics. https://
www.frontiersin.org/articles/10.3389/
frobt.2020.00036/full.
[6] Inman Harvey. The microbial genetic algorithm. In European con-
ference on artificial life, pages 126–133. Springer, 2009.
