1 of 60

FVD - Encore

Date: Wednesday, Nov 30 2022

2 of 60

Team H - TouRI

Prakhar Pradeep

Jigar Patel

Shruti Gangopadhyay

Jashkumar Diyora

Shivani Sivakumar

Software and�Interface Lead

Hardware and �Sensors Lead

Autonomous Navigation�Lead

Perception�Lead

Autonomous Manipulation Lead

3 of 60

TouRI Advisors/Sponsors

Zackory Erikson

Yonatan Bisk

RCHI Lab, RI, CMU

CLAW Lab, LTI, CMU

Assistant Professor

Assistant Professor

4 of 60

TouRI - Overview

5 of 60

Use Case

6 of 60

Use Case

7 of 60

Use Case

TouRI

8 of 60

Use Case

PAN-TILT DISPLAY

To provide human interaction

MANIPULATOR

To facilitate interaction with the environment

MOBILE BASE

To facilitate traversal

9 of 60

Use Case

10 of 60

Use Case

11 of 60

Use Case

12 of 60

Use Case

13 of 60

Use Case

14 of 60

Use Case

15 of 60

Autonomous Pick and Place

16 of 60

System Design

Receive user input from interface with a latency less than 5 seconds

17 of 60

System Design

18 of 60

System Design

19 of 60

System Design

20 of 60

System Design

21 of 60

System Design

22 of 60

Interface subsystem - Performance

Receive user input from interface with a latency less than 5 seconds

23 of 60

  • Jash shifts to autonomous manipulation mode

  • Robot sends an image to Jash with the detections

  • Robot autonomously picks and places the selected object in the shipping box

System & Demo Overview

24 of 60

Software Architecture - Autonomous Picking

25 of 60

Software Architecture - Autonomous Picking

PERCEPTION

MANIPULATION

26 of 60

Functional Architecture - Autonomous Placing

PERCEPTION

MANIPULATION

27 of 60

Automated data labelling pipeline

  • Annotate first image and track for the next images
  • 770+ images with 12 labels
  • Adding an additional object takes <30 mins

Dectectron2 framework

  • FasterRCNN Network FPN

Perception - 2D Pipeline

"0" : 'cmu_tartan_bottle',

"1" : 'tennis_ball_toy',

"2" : 'cmu_cup',

"3" : 'cmu_bottle',

"4" : 'monkey_keychain',

"5" : 'transparent_bottle',

"6" : 'all_star_dogs_belt',

"7" : 'dog_collar',

"8" : 'cow_keychain',

"9" : 'beanie',

"10" : 'unicorn',

"11" : 'airpods_case'

Labels

28 of 60

Souvenir picking 3D pipeline (C++)

3D Pipeline - Souvenir centroid estimation

Point Cloud Cropping

Centroid estimation

Souvenir Detection

29 of 60

Souvenir picking 3D pipeline (C++)

3D Pipeline - Shipping box centroid estimation

Point Cloud Cropping

Centroid estimation

Shipping box Detection

Plane detection

30 of 60

Autonomous pick and place

Performance Validation

31 of 60

Requirements Satisfied:

  • M.P.6: Detect objects with a precision of 70% and recall of 60%.

Subsystem Performance Results:

  • Object detection ( 12 Classes )
  • Precision - 81.43%
  • Recall - 94.36%

Perception subsystem - Performance

32 of 60

Manipulation subsystem - Performance

33 of 60

Average: 57.62 mins

Average: 58.87 mins

Manipulation subsystem - Performance

34 of 60

Requirement Metrics

Achieved Metrics

Detect objects with a precision of 70% and recall of 60%

Objects detected with a precision of 81% and recall of 94%

Estimate centroid objects with a precision of 65%

Centroid of objects estimated with a precision of 81%

Plan manipulator motion to grasp object within 3 minutes​

Manipulator motion to grasp object planned in 12 seconds

Grasp or place object within 5 minutes, with 67% success rate​ (2 successful trails out of 3)

Object grasped/dropped in 52 seconds, with 80% success rate (4 successful trials out of 5)

Manipulation and Perception Sub Systems - Performance

35 of 60

Natural Language Understanding and Cognition

36 of 60

Natural Language Understanding

Hey Touri, pick up a gift for my dog

WAKE WORD

SPEECH RECOGNITION

COGNITION

37 of 60

Natural Language Understanding

Multilingual speech recognition as well as speech translation and language identification

WAKE WORD

SPEECH RECOGNITION

COGNITION

38 of 60

Cognition

Fine-tuned GPT-3 (text-davinci-003)

OBJECT ID

Knowledge-base

Visual feedback

Command

39 of 60

Cognition

Fine-tuned GPT-3 (text-davinci-003)

OBJECT ID

Knowledge-base

Visual feedback

Command: Buy a gift for my dog

Object: Dog belt

40 of 60

Cognition

Fine-tuned GPT-3 (text-davinci-003)

OBJECT ID

Knowledge-base

Visual feedback

Command: My friend loves music. Pick a gift for him

Object: No object

41 of 60

Cognition

Command: Pick up the unicorn

Object: Unicorn

Command: Buy a gift for my dog

Object: Dog belt

Command: My friend loves music. Pick a gift for him

Object: Airpod case

42 of 60

Mechatronics Subsystem

and Interface

43 of 60

What you will see - Mechatronics/Interface

Provide upto 1080*720 resolution video call interface for user to interact with surroundings with a lag less than 2 seconds

44 of 60

What you will see - Mechatronics/Interface

Provide traversal feedback to user every 5 seconds​

45 of 60

What you will see - Mechatronics/Interface

Provide gimbal motion of 60 degrees in pitch and 120 degree in yaw for the display device​

46 of 60

What you will see - Mechatronics/Interface

Receive user input from interface with a latency less than 5 seconds

47 of 60

What you will see - Mechatronics/Interface

Receive user input from interface with a latency less than 5 seconds

48 of 60

Navigation Subsystem

49 of 60

Hard Case Test- Dynamic Environment

50 of 60

System & Demo Overview

2 modes of the navigation system- Autonomous mode and Teleoperation mode

Teleoperated Navigation

Autonomous Navigation

Tour location received from interface

Robot localization and path planning

Static and dynamic obstacle avoidance

Robot teleoperation through interface

Obstacle detection and user alerts through interface

51 of 60

Functional Architecture

Sensors

52 of 60

Functional Architecture

Robot Localization

53 of 60

Functional Architecture

Path planning & Obstacle Avoidance

54 of 60

Functional Architecture

Teleoperated control

55 of 60

Navigation subsystem - Functional Architecture

56 of 60

Navigation subsystem - Performance

Requirement Metrics

Achieved Metrics

Traverse on hard, flat indoor floors reliably at 0.4 m/s​ during teleoperation

Traversed on hard, flat indoor floors reliably at 0.4 m/s​ during teleoperation

Reach the desired location 50 m away within 30 minutes

Reached the location 60.96 m away within 5 minutes 34 seconds

Plan global path to the desired location within 3 minutes​

Planned global path to the desired location within 4.23 seconds​

Detect and avoid obstacles with mAP of 80%​ during autonomous navigation

Detected and avoided 90% of the obstacles

Detect obstacles with mAP of 80%​ during teleoperation

Detected 100% of the obstacles

57 of 60

Navigation subsystem - Performance

60 m autonomous traversal time

Time for path planning

Obstacle avoidance success

Average: 5.34 mins

Average: 4.23 secs

90% success rate

58 of 60

Thank You Everyone�

Special Thanks

Prof Zackory, Yonatan, John, Dimi and TAs

AI Makerspace - Greg

59 of 60

Time for Souvenir Giveaway!

Get your voucher from Jash and wait for next instructions

60 of 60