this is from IPG open house Shang Hai
scenario generation task
data record
tracking vehicles, roads, tobstacles
obj: lane, road, barries, GPS input, vehicle position/orientation, fixed ID, type
the goal of recoding is for road building, which will be used in replay.
road build
GPS input + lane mark info + vehicle location –> vehicle trajectory
replay
run config, input as tranversal and longitudial position
traffic vehicle location, speed
rearrange
input as : traffic vehile info + ego info , list of traffic vehicle info
traffic vehicle manage:
1) manuevor control: free move
2) spawn control: lati + longi --> 23 cases
3) support external plugins + manuevor trigger
Synthetic Scenario
junction assistant
road type + traffic rules + scenario –>
support road topology modification
support different envs: day of time, weather,
scenario editor to support opendrive import
standardization
PEGASUS + ASAM simulation standards
roads, scenarios, simulation interfaces
Opendrive –> road topology
opENScenario –> maneuver & anction abstract definitions
Open simulation interface –> interface developed for PEGASUS
limitations
pre-define route for vehicle ?
the ego car has AI maneuvor ?
Vitual Prototype
including gearbox loss mode, gas mode, through look-up table
including hybrid powertrain architectures: automatic gearbox + parallel hybrid
including powertrain masses(engine, tank, gearbox, battery, motor)
including trailer data set generator
including damping top mount
Simulation test
support:: ADAS/AD, POWERTRAIN, Vehicle Dynamics
steering system visual case
for less steering will overall comfort and vehicle dynamics
reference measurements(steering-in-loop simulator) -> model parameter id + softare + ECU integration –> parameterization & validation -> training
how the steering system works
open loop to get mechanical characteristics(stiffness, friction..)
system performance with or without EPS
1) ideal(basis) model vs physical model
how to cowok the physical model with autopilot control model ?
test bed
to support electrification, durability, balancing, driveability, powertain caillbratio, connected powertrain
AI training with synthetic scenario
decion making
trajectory planning
image perception
q: how to make sure AI robost ? –>
what CarMaker can do for AI?
1) obj annotation (vehicles, pedestrains) –> auto annotation
2) semantic segmentation
e.g. IPG Movier for auto semantic segmentation
Q: what’s the hardware for ?
Cloud & CPU/GPU for Parallelization
q: how to parallel in docker ?
1) test case in each CPUs
2) even for single test run(with multi sensors, multi cars )
resources & distribution
CPU: vehile model, drivel model, envs, ideal sensors
GPU; visual, camera RIS, radar ris, lidar rs
Test run in prallel
sensor setup(10 ultra, 5 Radar, 1 Lidar, 1 Camera)
host pc (with test manager) + 4 virtual machines
output: key figures, reports, statistics, queries
open archi for scalable processing( on-premise and cloud)
big data anaysis with DaSense by NorCom
- how it works ?
- external scheduler mananger, PBS
- HPC light to support local PC parallel
new features in 8.0
virtual test driving 8.0
- simulink lib (through Simscape)
- Scenario Editor: vege geenration, animated 3D objs, new models(vehicles, trailers, trucks, bus, buildings, houses, street furniture, pedestrains)
- visulize road surfaces ..
- ipg movie
- fisheye distortion from external file
- new sensor models(Lidar RSI)
q: what’s the difference of open source tool vs commericial ?
Lidar RSI
Ideal perfect world –> ground truth
HiFi –> false positives & negatives
raw data –> RSI
supporting Lidar type:
moving laser & photot diode
moving mirrors
solid state
flash
input features :
Laser beam, including custom beam pattern, Raytracing rays
Scene Interaction, including atmoshpere attenuation, color or material or surface or transparent dependency
detection, including threashold, multiple echoes per beam, separability
output features:
- sending & receiving direction of every beam
- light intensity of every beam
- time & lenght of light
- pluse width
- number of interactions
User Case : Nio Pilot
by sun peng
cases
inter-city, parking, closed space, crowded space
sensors: 3 front camera, 4 surround camera, 4 mm RADARS, 12 Ultra, 1 driver monitor camera
higway pilot in June
perception: camera, radar, ult, hd map, location
planning : path planning, maneuvor decsion, system control
cloud & AI
simulation usage
FDS -> cases -> SIL
platform –> cases -> regression test, abstraction & instantiation ; scene reconsturction(in-house) / close loop SIL ; traffic model training(to do)
integration -> HIL
what about vd ? –> co-work with simulation and physical test, the cover percentage of simulation is about 80%, the left is from
data platform
upload nodes -> cloud
med usa API server -> fleet mgmt
log stash –> elastic search –> Kibana & Admin (tensn and spark )
I think they are collecting data, and this data for scene building and simulation usage in future
data visulazation
HIL
lane model simualtion on HIL
fusion simulation on HIL
automation test
jenkins master –> jenkins slave (agent IPG) –> cloud
goal
simulation server <—> data center
parallel sim core + simulation monitor (data exchange service)
data processing + labelling + case management + traffic model training
(replay, SIL, REMODEL, Visuliazation )