ROS, German, CARLA, Open Source Simulation for Driving, Intel, June 29, 2018

<p class="MsoNormal" >ROS,&nbsp; German,&nbsp; CARLA, Open Source Simulation for Driving,&nbsp; Intel,&nbsp; June 29, 2018

This is a noon talk at Stanford in the Comp. Sci Dept in Gates 104.

German&nbsp; (pron.&nbsp; Herman; from Barcelona)&nbsp; works at Intel.

The CARLA research group is largely in Barcelona.&nbsp; (14 researchers.)

CARLA&nbsp; open source simulation for autonomous driving;

based in <b>Unreal Engine 4;</b>

modular, scalable, and flexible design...

Goal:&nbsp; democratizaiton of Autonom&nbsp; driving&nbsp; R and D

and standardized&nbsp; platform for AD = auton driving&nbsp; R and D assessment.

he shows a video...

with busy traffic;&nbsp;&nbsp;&nbsp; realistic uban environment...

lots of pedestrians;&nbsp; diverse weather;&nbsp; lighting conditions...

dozen in project...

different towns, different weather,&nbsp; lighting;

straight driving versus one turn&nbsp; vs navigation vs&nbsp; navig dynamics;

with a planner&nbsp; vs&nbsp; IL vs&nbsp; reinforcement learning (hardest)

support from Intel and from Toyota/ SAIL.

RGB images,&nbsp;&nbsp; range finding,&nbsp; bounding boxes;

training&nbsp; includes&nbsp;&nbsp; very severe rain;

see eg&nbsp; End to End&nbsp; via&nbsp; Conditional Imitation Learning...

Conditional Affordance ....

keeping steady distance;&nbsp; slow down appropriately;

stop for pedestrian;

*

simulation ot real&nbsp;&nbsp; policy transference..

he shows a video&nbsp; of&nbsp;&nbsp; a (real) toy car&nbsp; (about&nbsp; 1 meter in length)

that's been trained via&nbsp;&nbsp; CARLA.

LYFT Perception Challenge&nbsp; (Udacity)

CARLA was used by several participants.

CARLA Road Map&nbsp; (for future dev.)

Feature Requests...

main lines of dev.

Sensors,&nbsp; Integration;&nbsp; Asset Library;

milestone-driven dev.

OSVF&nbsp;&nbsp; Open Source Vision Foundation&nbsp;&nbsp; (parent org)

goal:&nbsp;&nbsp; agents learn to drive safely in&nbsp;&nbsp; eg&nbsp; Boston,&nbsp; Palo Alto

in critical scenarios not seen before;

procedure generated scenes based on real locations

critical&nbsp; scenarios from NHSTA pre-crash reports...

CORE:

control all agents of the simulation

eg&nbsp;&nbsp; vehicles,&nbsp; pedestrians,&nbsp; traffic lights;

want automatic asset loading&nbsp; programmatically.

want to run automatically

want advanced physics and terrain simulation

performance improvement....&nbsp;&nbsp; in terms of frames per second...

automatic content generation...

CityEngine &nbsp;content generation...

eg&nbsp; trees, buildings;

improve sensors:&nbsp; LIDAR,&nbsp;&nbsp; cameras&nbsp;&nbsp; etc.

integrate with ROS,&nbsp; SUMO, etc.

join our&nbsp; DISCORD community...

14 engineers in his group;

Qs&nbsp;&nbsp; about multilane&nbsp; and about freeway overpasses;

&nbsp;

&nbsp;</p>