Autonomous Ground Systems ADVANCED PERFORMANCE COMPUTING POWER FOR - - PowerPoint PPT Presentation

Autonomous Ground Systems

8/9/2018

ADVANCED PERFORMANCE COMPUTING POWER FOR LARGE DATA SET PROCESSING

Karl Rosenberger, RAVE Sara Blackmer, RAVE Steve Wesoloski, RAVE John Brabbs, TARDEC

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

• Background/Problem - Challenges of real world robotic

and autonomous systems testing

• Hypothesis - Virtual testing provides complement to

real world testing

• Solution

– Simulation Environment – Processing Hardware

• Key Take-aways/Conclusion
• Future

Overview

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

Robotic & Autonomous Systems (RAS) Leader-Follower (LF) System Concept

• Uses Large Unmanned Ground System (LUGS) - Manned lead vehicle followed by unmanned

followers

• Uses an Applique architecture for the current manned vehicle fleet to automate the vehicle

via three different kits: – Autonomy kit (Akit) - Composed of perception sensors and computers for processing sensor data for implementing high level autonomy behaviors – By-Wire kit (Bkit) - Permits control of all primary vehicle controls and provides feedback and control of platform components – Warfighter Machine Interface (WMI)/Radio Network Interface (RNI) kit - Allows the communication and control of the convoy (Ckit)

• These kits will be used to retrofit existing vehicles with autonomous capabilities.

Background/Problem

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

Challenges of real world robotic and autonomous systems (RAS) testing: “Before being fielded, Army RAS will need to be rigorously tested under representative operational conditions by soldiers in the unique environments where the RAS will operate.” – Safety – Hardware and Software failure potentials – Test facility limitations – Terrain, weather, traffic, event scenarios… – Driver availability – 1 per vehicle (minimum) – Testing equipment/trucks – March Unit consists of 2 to 4 vehicles – fully instrumented – Time required – 1 million miles = 3.42 yrs (50mph average for 16 hrs/day) – Cost - $$$

Background/Problem

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

• Virtual-World RAS Testing could provide a vital complement

to real world or live RAS testing. – Can a simulation environment overcome or reduce challenges posed in real world testing?

• Can a virtual simulation environment (software) be developed

that replicates real world conditions?

• Can a physical simulation

environment (computer solution) be developed to support this virtual world?

Hypothesis

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

Continuous Autonomy Simulation Test Laboratory Environment (CASTLE)

– Simulation Environment (Virtual) – Processing (Hardware)

Solution

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

TARDEC RAS M&S Capability

• Enhance current real-time M&S capability

– To evaluate and test edge cases – Augment not replace real world testing – Use controllability, repeatability, and scalability available in virtual world

Solution – Simulation Environment

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

CASTLE HW Configuration LF System Concept (2 vehicles)

• Each RAVE HPC workstation in CASTLE represents one vehicle.
• This is an improvement over the workstations that TARDEC used on a previous project that required two

workstations per vehicle, one to represent the virtual simulation environment, one to support the autonomy software.

• CASTLE can currently support up to an eight vehicle convoy, two four vehicle convoys, or four two vehicle

convoys (shown below).

Solution – Processing Hardware

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

Customized configuration required

• No suitable solution available off the shelf
• Each component of system must be best of breed

RAVE HPC Workstation Specifications:

Solution – Processing Hardware

Motherboard Intel X99 Chipset Processor Intel Xeon multi-core, multi-thread (14 cores, 28 threads) RAM 64 GB GPU GeForce GTX 1080Ti

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

Key aspects of system design

– Compact form factor

• Short-depth 4U rackmount chassis

– Reliable

• Server-grade chassis, cooling and components
• Custom cooling system installed, high airflow
• IPMI for remote management

– Optimized

• Started with the most ideal motherboard on the

market – Low latency, high throughput – Upgradeable

Solution – Processing Hardware

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

1. Data captured from real-world testing of concept being used

• CASTLE replay’s the leader vehicle positions
• The followers’ autonomy software execute in real time in the virtual simulation environment.
• Location of the follower during the real test is replayed as a ghost vehicle to compare with the

simulated follower.

• The WMI messages are also replayed to change the gap distance

Key Take-aways

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

2. Dynamic obstacles between leader and follower

• The test below shows how moving a dynamic obstacle in between the leader and the follower

verifies that the follower will stop for the dynamic obstacle

• This provides evidence that the simulated LIDAR is providing the correct input to the autonomy

software to recognize an obstacle in its path.

Key Take-aways

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

2. Dynamic obstacles (continued)

• Once the dynamic obstacle moves out of its path of the leader vehicle, follower resumes
• Again this provides evidence that the simulated LIDAR is providing the correct input to the autonomy

software to recognize no obstacle in its path.

Key Take-aways

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

3. Enables thousands of iterations by only changing a few parameters:

– Angle of the sun, weather, dynamic and static obstacles – LF gap distance and convoy speed

Key Take-aways

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

REAL-WORLD TESTING (PROBLEM) VIRTUAL-WORLD TESTING (SOLUTION) Test Range 2 test range personnel required No range personnel needed 1 test engineer required 1 test engineer to setup the scenarios to run Weather conditions are limited to those available at the time and location of testing (e.g., if you want to do cold weather testing than you will need to go to a cold weather region) Computer simulation can easily model different weather conditions and times of day Terrain type is limited to what is available at the test range Infinite terrain types are possible Test range availability is limited Virtual world is available 24/7 Test range usage has a cost per hour Once virtual terrain is built, it can be used as many times as needed with no additional cost Drivers Lead vehicle = 1 driver Lead driver is automated Safety drivers for Followers (e.g., 3 Followers = 3 people) No safety drivers are necessary

Key Take-aways

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018 DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Key Take-aways

REAL-WORLD TESTING (PROBLEM) VIRTUAL-WORLD TESTING (SOLUTION) Test Vehicles Testing a single March Unit (1 Leader + 3 Follower vehicles) requires four vehicles Trucks are all virtual so there is no limit Simulating a convoy comprised of multiple March Units requires four more physical vehicles per additional March Unit Simply add more computers if more trucks are needed 1 - 2 spare vehicles are needed on site to fill in for maintenance or fuel interruption, preventing a delayed

• r degraded test as a result of

temporarily missing vehicles No spare vehicles needed, virtual vehicles don't break down or need to refuel Additional trucks may not be available; Each truck adds an

• ngoing cost

There is a very low one-time cost for an additional computer (i.e., truck) Fuel cost is ongoing for all testing No fuel cost Vehicle maintenance cost (mechanic available) No vehicle maintenance cost

Autonomous Ground Systems

8/9/2018

REAL-WORLD TESTING (PROBLEM) VIRTUAL-WORLD TESTING (SOLUTION) Obstacles Physical obstacles must be available and are vulnerable to damage One time cost for creation of virtual

• bstacles which can then be used

as many times as needed Could damage vehicle Represents traffic, crowds and

• bstacles virtually

Degraded Modes Sensor failure - hard to represent in real world Can be scripted to turn off sensor in software

Key Take-aways

CONCLUSION:

As demonstrated with the PdM ALUGS CASTLE system, HPC workstations and M&S capabilities can be combined to effectively and efficiently test RAS in a cost-effective and timely way.

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

HPC with GPU compute

• Exponential growth of data requires rapid

development of HPC hardware and software

• Future HPC efforts require different

technologies

– Traditionally workload spanned across a cluster

• f CPU machines

– Now, parallel processing across GPUs is more economical and efficient

• 5120 cores vs 32 cores

Future

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

What about AI opportunities?

Future

2025? 2010

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

AI opportunities

• Comb through data quickly and produce actionable information
• AI can predict behavior and identify edge cases where real-world driving may never isolate
• r may be too dangerous to test

Future

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.

Autonomous Ground Systems

8/9/2018

Thank you

DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited.