AI Think I Can: Why GPUs Could Be the Engine Leading Us to Autonomous Trains

Boosted by machine learning, image recognition and NVIDIA GPUs, trains are on track to lead the way in autonomous transportation.
by Scott Martin

Self-driving cars and trucks are the talk of the town, but trains could be first out of the station as a fully autonomous form of transportation.

That’s because — unlike passenger vehicles or long-haul trucks — a train moves on tracks, limiting the environment in which it needs to understand its surroundings and simplifying the types of decisions that need to be made.

Train systems have been built on decades-old technologies — from hardware-based signal systems to radio frequency-driven dispatchers.

Now the professionals who make the trains run on time are shifting to tools such as the internet, sophisticated sensors and, increasingly, GPU-powered deep learning.

That’s a big change for an industry that was a hallmark of the original industrial revolution.

‘Future of Smart Trains’

“We’re going into the future of smart trains and dumb tracks. Technology is changing from track technology to in-train technology,” said Derel Wust, managing director at 4TEL Pty Ltd, an Australian private company working on a deep learning pilot program in New South Wales.

An example of the high costs for older train technology is playing out in New York’s multibillion-dollar MTA communications-based train control subway project, Wust said in a talk at the GPU Technology Conference earlier this year. Governor Andrew Cuomo said in a speech that at this rate it will take 40 years to upgrade the entire system.

A report estimated the MTA project would cost about $20 billion over a 35-year period. Wust said that a train-based AI solution would minimize a big portion of that expense and be completed quicker.

4TEL is focused on improved safety, reduced infrastructure and continuous machine learning for train systems. Many see the role played by train operators improved by automated acceleration and braking systems. “Many of the accidents we see in our daily lives can usually be attributed to human error,” Wust said.

Working with John Holland, an engineering company, 4TEL has a contract that provides the opportunity to bring its new machine learning-based approach to trains on the Country Regional Network of New South Wales. At this stage, Wust wants to collect data and see how it works with behavioral models.

The company has a pilot program in New South Wales for its Horus box, which uses infrared and optical cameras and machine learning algorithms to provide a software service for train operation. The box runs on the NVIDIA DRIVE platform.

Tracking Simulations

Wust’s company is among a handful of railway firms pioneering autonomous trains using so-called “digital twin” models, which allow simulations on tracks that mirror their rail networks.

Swiss Federal Railway Company, which manages nearly 5,000 miles of railways in a country that’s 200 miles wide, was at GTC to talk about its efforts to update train systems to keep them running on time.

It maintains a complex system that comes from a combination of high-speed trains and slow-moving freight trains, requiring some 13,000 switches to help control its mixed train traffic.

Known as SBB, the company shuttles more than 1.2 million passengers a day. It operates its own power plants that supplies its trains almost exclusively with renewable energy and is one of Switzerland’s largest real estate companies. “What might surprise you is that we are also a big software company,” said Erik Nygren, a business analysis and AI researcher at the company.

The company’s Research and Innovation Platform, led by Dirk Abels, uses the NVIDIA DGX-1 AI supercomputer for simulations as well as for deep reinforcement learning to optimize train schedules and dispatching.

SBB has integrated all of its geographic information into its simulation environment, allowing for interactions with real-time train data by train dispatchers in the virtual setting.

Also, it has developed safety features to give locomotives automatic braking systems, which are currently working in simulation. “We have collision detection, which is able to operate within a half second,” said Adrian Egli, a business analyst and HPC expert for SBB.

Maintenance Tracking

In addition, General Electric is working on in-train systems that integrate cameras, software and GPU technology for guiding trains. GE has a contract with Indian Railways for a locomotive in Bangalore, India, to pilot its technology.

The company sells the equipment as well its subscription-based camera system that includes analytics. GE Transportation uses the NVIDIA DGX-1 as well as other NVIDIA GPUs for training. For inferencing of models on-board the locomotive, GE is looking at the NVIDIA Xavier platform and P100 GPUs.

Inspections can be an ordeal that shuts down segments of tracks while people walk them looking for faulty railroad ties and other maintenance issues, costing millions of dollars in down time. GE Transportation is working to change scheduled maintenance to “predictive and prescriptive” using AI, saving money without compromising safety.

GE is working on using its system of front-facing cameras for train safety as well as track inspection. “Can we make the locomotive smart enough and self-aware to start inspecting its own track?” said Dattaraj Rao, a principal architect at GE Transportation, who spoke at GTC.

Like others, GE is taking a so-called digital twin approach, or simulated railway network. It is using it for constant insights on track conditions and to help schedule maintenance.

4TEL’s Wust said it’s pretty easy to imagine that trains gathering data via onboard cameras and sensors could collect data from opposing tracks, such as a rail obstruction. This could provide Waze-like traffic information onto the railway system for other trains.

“We started investigating these images to try to extract value from them,” Rao said.