KiwiBots: Using the power of GPUs to solve the last mile delivery - - PowerPoint PPT Presentation

▶
kiwibots using the
SMART_READER_LITE
LIVE PREVIEW

KiwiBots: Using the power of GPUs to solve the last mile delivery - - PowerPoint PPT Presentation

Feb 26, 2024 371 likes •907 views

KiwiBots: Using the power of GPUs to solve the last mile delivery problem $whoami My name is David Cardozo Machine Learning Engineer at Kiwi Campus. I am: @davidcardozo Linkedin @_davidcardozo twitter Juan Galvis Carlos Alvarez AI Lead

slide-1
SLIDE 1

KiwiBots: Using the power of GPUs to solve the last mile delivery problem

slide-2
SLIDE 2

$whoami

My name is David Cardozo

Machine Learning Engineer at Kiwi Campus. I am: @davidcardozo Linkedin @_davidcardozo twitter

slide-3
SLIDE 3

AI Lead - Kiwi Campus Email: charlie@kiwicampus.com LinkedIn: @calvarez92

Carlos Alvarez Juan Galvis

Robotics Lead - Kiwi Campus Email: juangalvis@kiwicampus.com LinkedIn: @jgalvis-mechatronics

slide-4
SLIDE 4

Computer Vision Engineer Email: john@kiwicampus.com

John Betancourt Andres Rengifo

Hardware Lead Email: andres@kiwicampus.co m LinkedIn: @andresr8

Juan Jurado

Electronics Engineer Email: jj@kiwicampus.com LinkedIn: @juanfjuradop

slide-5
SLIDE 5
  • 1. Kiwi in LatAm market.
  • 2. The Last Mile Delivery Problem
  • 3. How did we end-up in robotics.
  • 4. Computer Vision
  • 5. Convolutional Networks for Driving, and Image

tasks

  • 6. Jetson TX2
  • 7. Development
  • 8. KiwiBot

Agenda

slide-6
SLIDE 6

1.

Kiwi Campus? How to make deliveries in the United States?

slide-7
SLIDE 7
slide-8
SLIDE 8
slide-9
SLIDE 9

PLAN

How to make deliveries < $1

slide-10
SLIDE 10
slide-11
SLIDE 11
slide-12
SLIDE 12

$25

Per person

< 3 orders per hour

Extremely difficult to get to 15 deliveries per hour

< 30

Available students

slide-13
SLIDE 13

“

“That’s the hard thing about hard things—there is no formula for dealing with them.”

slide-14
SLIDE 14

$25

Per person

< 3 orders per hour

Extremely difficult to get to 15 deliveries per hour

< 30

Available students

slide-15
SLIDE 15

◎ Less that one dollar delivery. ◎ Almost 40% of the cost of a delivery is on the last mile delivery. ◎ Robots can help us out to minimize costs and time.

The last mile delivery problem

slide-16
SLIDE 16

Robots for delivery?

slide-17
SLIDE 17

3.

How did we end up in robotics?

It is not rocket science…..

slide-18
SLIDE 18
slide-19
SLIDE 19
slide-20
SLIDE 20
slide-21
SLIDE 21

The age of Deep Learning

  • 1. Autonomy for

navigation and delivery

slide-22
SLIDE 22
slide-23
SLIDE 23

“

“So how exactly the computer sees? – The thing is most of computer vision researchers do not really understand how the computers see. It’s like alchemy and chemistry. Alchemy came first and chemistry came then. And right now we are in the alchemy stage of computer vision, where it works but we are not sure

  • why. And it is the chemistry stage that I look forward to.”
  • Bill Freeman

Computer Vision

slide-24
SLIDE 24

Statistics of natural images obey invariants

… Translation Cutout Dilatation Contrast Rotation Scale Brightness …

slide-25
SLIDE 25

Invariant under Translation ---> Convolution (Cross-Relation)

slide-26
SLIDE 26

The visual pathway

slide-27
SLIDE 27
slide-28
SLIDE 28
slide-29
SLIDE 29
slide-30
SLIDE 30
slide-31
SLIDE 31
slide-32
SLIDE 32
slide-33
SLIDE 33
slide-34
SLIDE 34

Camera Calibration

slide-35
SLIDE 35
slide-36
SLIDE 36
slide-37
SLIDE 37
slide-38
SLIDE 38

Data Curation

slide-39
SLIDE 39
slide-40
SLIDE 40
slide-41
SLIDE 41
slide-42
SLIDE 42
slide-43
SLIDE 43

Hello!

I am a Kiwibot

◎ Low cost ◎ Navigation using machine learning. ◎ Currently serving students in the Bay Area.

slide-44
SLIDE 44
slide-45
SLIDE 45

t HTML CSS

Technology Stack

Tensor Flow JavaScript Flask AI Jupyter NodeJS SocketIO JSON Java ROS Pandas Numpy REST Python C++ cv2 FloydHub Docker

The AI Iceberg

CUDA TensorR T

slide-46
SLIDE 46

The iceberg of research

t Pilot-net

Teaching a Robot to drive.

Object Detection:

Image Parsing, also Image segmentation Forefront of Kiwi’s Data Science Team

Final product: End-to-end delivery at low cost Simulation

Domain Randomization: Real data is expensive, way too expensive

Reinforcement Learning

Be better that the human pilot.

Crossing Detection

Image Classification

slide-47
SLIDE 47

What’s Next

◎Temporal Analysis ◎Predicting Throttle (velocity) ◎Multitask Learning ◎Feed GIS information to the NN ◎Transfer Learning from Virtual Environments

slide-48
SLIDE 48

Deployment

ResinOS

◎Linux containers for IoT ◎Lightweight payloads ◎Brick-safe deployments ◎Phased deployments, scheduled updates ◎Device status, location, deployment logs - > all on an ongoing basis

slide-49
SLIDE 49

Resin OS

slide-50
SLIDE 50

Resin OS

slide-51
SLIDE 51

Example

slide-52
SLIDE 52

We are hiring!

Contact:

david@kiwicampus.com f@kiwicampus.com

slide-53
SLIDE 53

∙

The future today