System Architectures Basics

System Architectures

Basics

Duckietown

System decomposition

• All starts with a decomposition of a system in components.

• What is a system and what are components?

2

“A system is composed of components;a component is something you understand.”

— Howard Aiken (1900-1973)

Duckietown

Physical and logical architectures

• There are two aspects for “architectures”:

• Logical architecture - what the system is doing

• e.g., system decomposition, data flow

• Physical architecture - how it is doing it

• e.g., which computer runs which component

• Containerization: Which computer runs which virtual computer that runs which component?

3

Duckietown

Logical architecture

• The logical architecture describes:

• System decomposition in components

• Data flow (Who tells whom what)

• Representations

• Priors (Who knows what)

• The logical architecture is independent of language, middleware, and other implementation details.

4

Something Useful

Duckietown

control

5

given

to design

inputsoutputs

process and environment

agent / controller

world / plant

belief / estimate

perceptionplan

planning

A familiar example (autonomy) logical architecture

commands

actuators

observations

sensor(s)

Duckietown 6

Modeling

𝑐

𝑥𝐴

𝑦𝐼

𝑥𝐼

𝑦𝑊

𝑥𝑊

𝐴

𝐶

𝑣𝑙

𝑣𝑟

𝑂

I𝑑

𝑦𝐴

2𝑅

𝜔

2𝐿

𝜃

𝜑𝑙

𝜑𝑟

𝑦𝑅𝑥𝑅

Duckietown 7

Sensing

Z

YX

CameraCenterC

PPrincipalPoint

ImageP

lane

PrincipalAxis

u Aa

v

v

y:;<

x:;<

Duckietown 8

Planning

Duckietown 9

Control

Duckietown 10

Perception

https://www.youtube.com/watch?v=OksuVuNY5o0

Duckietown

Part of logical architecture for Duckietown• Part of the main perception pipeline.

• Purple refers to “static” information.

• The diagram does not show how things are implemented.

• It does show who-knows-what, and who-tells-what-to-whom.

11

sensor

learned transform

image anti-instagram

anti-instagram learner

transformed image

segment detector

ground projection

image-spacesegments

appearance prior

calibration

calibration procedure

3D segments

Duckietown

Computation graph

• Computation graph

• nodes = components

• edges = signals

• node weight = flops required

• edge weight = size in bytes

12

Duckietown

Physical Architecture

• “Physical” architecture: how things are implemented.

• The physical architecture includes:

• Which processor runs what process?

• How is the data communicated (TCP, UDP, etc.)

• Where is the data stored

• Protocols, formats, etc.

13

Duckietown

Physical architecture: Resource graph

• Resource graph:

• nodes = processors

• edges = network links

• node weight = processor power

• edge weight = bandwidth

14

Duckietown

Part of physical architecture of Duckietown• This is the Duckietown ROS graph

• It lists the nodes and the signals.

15

legend Sensors

Camera

Actuators

Visualization for operator

Line Detection and Projection

Parallel AutonomyFinate State Machine

Intersection Navigation

Coordination

F1: Illumination Robustness

F2-3: LED Detection

F4: Odometry Calibration from Sensor Measurements

F5: Lane Filter and Control

F6: Vehicle Detection

F8: Apriltags

F9: MDOAP

missing

designed

development

compiled

documented

tested

pi_cameracamera_node (SY,AC)

~image/compressedsensor_msgs/CompressedImage.msg

joy_node

joysensor_msgs/Joy.msg

joy_mapperjoy_mapper_node (SY)

led_joy_mapperled_joy_mapper_node

pi_cameradecoder_node (SY)

pi_cameracam_info_reader_node (SY)

line_detectorline_detector_node (LP,HZ)

anti_instagramanti_instagram_node (F1)

led_detectionLED_detector_node

camera_node/image/rawsensor_msgs/Image.msg

camera_node/camera_infosensor_msgs/CameraInfo.msg

ground_projectionground_projection_node (CC)

veh_detectorvehicle_filter_node (F6)

apriltags apriltags_global_node (F8)

veh_detectorvehicle_detector_node (F6)

apriltags april_pre_pros_node (F8)

mdoapstatic_object_detector_node (F9)

calibration/camera_intrinsic

led_interpreterLED_interpreter_node

led_emitterled_emitter_node

dagu_carwheels_driver_node (DY,SY)

wheels_driver_node/wheels_cmdduckietown_msgs/WheelsCmdStamped.msg

kinematicsforward_kinematics_node (F4)

~change_color_patternString

~image_with_linessensor_msgs/Image.msg

~object_detection_markersvisualization_msgs/MarkerArray.msg

~segment_listduckietown_msgs/SegmentList.msg

~segment_list_projduckietown_msgs/SegmentList.msg

visual_odometryodometry_training_pairs_node (F4)

lane_filterlane_filter_node (LP)

stop_line_filterstop_line_filter_node (LP)

calibration/camera_extrinsic

~joy_controlduckietown_msgs/Twist2DStamped.msg

lane_controllane_supervisor_node (LP)

~lane_control_supervisedduckietown_msgs/Twist2DStamped.msg

dagu_carcar_cmd_switch (SY)

fsmfsm_node (MN)

~modeduckietown_msgs/FSMState.msg

~car_command_switchduckietown_msgs/Twist2DStamped.msg

intersection_controlopen_loop_intersection_control_node (MN)

veh_coordinator_node (DH)

~intersection_controlduckietown_msgs/Twist2DStamped.msg

~intersection_doneduckietown_msgs/BoolStamped.msg

navigationrandom_april_tag_turns_node (MN)

~turn_typestd_msgs/Int16.msg

navigationgraph_planner_node (MN)

navigationaction_dispatcher_node (MN)

navigation/GraphSearch.srv


~intersection_goduckietown_msgs/BoolStamped.msg

~corrected_image/rawsensor_msgs/Image.msg

~image_healthduckietown_msgs/AntiInstagramHealth.msg

~raw_led_detectionled_detection/LEDDetectionArray.msg

~signals_detectionled_detection/SignalsDetection.msg

~desired_signaling_statestd_msgs/String.msg odometry_training_pairs

kinematicskinematics_learning_node (F4)

calibration/inverse_kinematicscalibration/forward_kinematics

kinematicsinverse_kinematics_node (F4)

~poseduckietown_msgs/Pose2DStamped.msg

~velocityduckietown_msgs/Twist2DStamped.msg

~lane_poseduckietown_msgs/LanePose.msg

~in_laneduckietown_msgs/BoolStamped.msg

~at_stop_lineduckietown_msgs/BoolStamped.msg

~stop_line_diststd_msgs/Float32.msg

lane_controllane_controller_node (SY,SC)

~lane_controlduckietown_msgs/Twist2DStamped.msg

~cornersduckietown_msgs/ObstacleCorners.msg

~vehicle_poseduckietown_msgs/VehiclePose.msg

veh_detectorvehicle_avoidance_control_node (F6)

~controlduckietown_msgs/Twist2DStamped.msg

~flagduckietown_msgs/BoolStamped.msg

~fast/image_raw sensor_msgs/Image.msg

~global/image_raw sensor_msgs/Image.msg

apriltags apriltags_fast_node (F8)

~tags_image sensor_msgs/Image.msg

~apriltags duckietown_msgs/apriltags.msg



apriltags apriltags_postprocessing_fast_node (F8)


apriltags apriltags_postprocessing_global_node (F8)


apriltagsDB.yaml

~detection_listduckietown_msgs/ObjectImageDetectionList.msg

mdoapobstacle_safety_node (F9)

~object_too_closeduckietown_msgs/BoolStamped.msg

mdoapsimple_stop_controller_node (F9)

~detection_list_projduckietown_msgs/ObjectProjectedDetectionList.msg


legend Sensors

Camera

Actuators





Coordination


F2-3: LED Detection




F8: Apriltags

F9: MDOAP

missing

designed

development

compiled

documented

tested



joy_node




































fsmfsm_node (MN)














































apriltagsDB.yaml







legend Sensors

Camera

Actuators





Coordination


F2-3: LED Detection




F8: Apriltags

F9: MDOAP

missing

designed

development

compiled

documented

tested



joy_node




































fsmfsm_node (MN)














































apriltagsDB.yaml







Duckietown

Deployment: Mapping logical architecture onto physical

• We need to map the computation graph onto the resource graph.

• Different choices will have different properties for latency, frequency, etc.

16

computation graph resource graphoptimization

deployment

Duckietown

• We need to map the computation graph onto the resource graph.

• Different choices will have different properties for latency, frequency, etc.

17

computation graph resource graph

deployment

optimization

Deployment: Mapping logical architecture onto physical

Duckietown

Example in Duckietown: Computation graph

• Computation graph: we collapse all computation in one node

18

camera driver

motors driver

(abstracting all computation as one node)

Duckietown

Example in Duckietown: Resource graph

19

camera actuator

4 core ARM

GPU

Raspberry PIYour laptop

GPU

4 core x86

wireless link

latency: 2 ms - 20 msbandwidth: 1-100 MB/s

Duckietown

Example in Duckietown: Deployment 1

• Option 1: Run everything on the Raspberry PI

20

camera actuator


camera driver

motors driver

orchestrator

Duckietown


• Option 2: Run from your laptop

21

camera actuator


camera driver

motors driver

orchestrator

Duckietown


• Option 3: Run heavy processing on the laptop

22

camera actuator


camera driver

motors driver

orchestrator

Duckietown


23Duckietown

How is all that implemented in Duckietown?

• Dozens of nodes in Duckietown, hundreds of connections

• Multiple different deployment scenarios:

• On a Duckiebot

• In simulation

• As an AI Driving Olympics (AIDO) submission

• We keep it clean and manageable by splitting the functionality in multiple containers

Duckietown


24Duckietown

The Docker images family tree

Duckietown 25Duckietown

Example deployments

Driving with a joystick Driving through the Dashboard

Running a demo

Evaluating AIDO submission in

simulation

Running a demo in simulation

Evaluating AIDO submission on a

Duckiebot

An interesting read

• E. A. Lee and S. A. Seshia, Embedded Systems -- A Cyber-Physical Systems Approach

System Architectures Basics

Documents