Story Highlights

• Indian AI researchers have built the world’s first open-source, fully-controllable multi-agent driving simulator.
• The simulator has the ability to model a multi-vehicle environment closer to their real-world behaviour in different traffic scenarios.
• Support for multi-agent algorithms will enable the creation of more realistic and reactive autonomous driving behaviours, the researchers say.

Researchers testing and tweaking machine learning algorithms used to control fully autonomous or self-driving vehicles have for long had to rely simulators that mimic real-life driving conditions while testing code that powers autonomous vehicles. A pedestrian crossing the road, a traffic signal turning from green to amber, a car ahead signaling a right turn, a school zone, and the hundreds of other scenarios in daily life on roads.

But, as much progress as these simulators have helped researchers – so-called Level 5 vehicles or those that need no human intervention, predicted to go to market as early as 2019 – make, they have a wall or another around them: they are either proprietary, inaccessible products like those in use at Uber and Google or they simulate conditions for single-vehicle conditions on proprietary platforms (e.g.:Carcraft). Even open source simulators – think Carla and DeepDrive as instances – control algorithms one vehicle at a time.

That’s set to change courtesy of researchers from two Indian Institutes of Technology: IIT-Madras and IIT-Kharagpur.

Released on Github last month, MADRaS (Multi-Agent DRiving Simulator) is a multi-agent version of TORCS and it adds the ability to use control algorithms on multiple cars running simultaneously on a track. The researchers say it’s the world’s first open source, fully-controllable, multi-agent driving simulator. It’s been created chiefly by Abhishek Naik, a 22-year-old from IIT Madras, while doing an internship at the Parallel Computing Lab at Intel India (PCL-India) in Bengaluru.

“When researching the open source world, I identified a couple of hot burning fires that needed to be doused,” teammate Anirban Santara told FactorDaily on a conference call. Santara, a 25-year-old Google India Ph.D. Fellow from IIT-Kharagpur, identified this particular deficiency in open source simulators while also interning at the PCL-India lab under Bharat Kaul, its director. “The entire thing was written by Abhishek ,” Santara says.

Balaraman Ravindran, head of Robert Bosch Centre for Data Science and Artificial Intelligence at IIT-Madras provided additional expertise in the areas of reinforcement learning. (Reinforcement learning is a branch of machine learning which provides a powerful learning paradigm to go beyond human capabilities. Google’s DeepMind, which defeated world champions at the game of Go, used reinforcement learning techniques, for example.)

A World First

As an AI researcher with exposure to problems related to fully autonomous driving, Naik says that most open-source driving simulators ( Carla, DeepDrive, and Airsim included) support control algorithms for a single car even if they come with pre-programmed behaviours of the other vehicles in the testing environment.

“Why is this problematic? If you want to simulate the traffic congestion in front of K R Puram railway station in Bengaluru, for example, these simulators would not be able to cater to the task,” says Naik. “A single agent has to learn to negotiate all types of real-world scenarios all alone, even though there are hundreds of vehicles around, each trying to achieve the same objective of reaching safely and reliably from Point A to Point B. It restricts the diversity of real-world scenarios that can be simulated.”

MADRaS allows each car on the driving track to be independently controlled in custom-made traffic scenarios.

“Today’s open-source multi-agent simulators for driving, to the best of our knowledge, require proficiency in niche and heavy low-level software like Unreal Engine or ROS (Robot Operating System), something that rarely appears in the skillset of a machine learning engineer or scientist,” says Santara. The major players in the autonomous driving space – Google or Uber, for instance – have their own in-house simulators and almost all of their software is proprietary, he adds. “The absence of an open-source multi-agent driving simulator has left the machine learning community throttled for a long time,” adds Santara. “We really wanted to build a basic set of tools that would let anyone to try their hands out and test the feasibility at a low cost.”

MADRaS is that attempt to lower the bar of entry for researchers in autonomous driving, says Ravindran. In comparison, “some of the more detailed simulators would require a steeper learning curve.”

Low specs, wide reach

It has minimal hardware requirements, says Naik, adding there’s no need for a GPU. “It even works on a five-year-old Core i3 laptop,” he says. “You can create your own traffic environment and assign custom behaviours to your cars, add as many cars as you want.” This potentially opens up research in multi-agent reinforcement learning and imitation learning research aimed at acquiring human-like negotiation skills in complicated traffic situations. “It’s a major challenge in autonomous driving that all major players are racing to solve,” Naik adds.

The goal was to build a platform on which you can quickly try out ideas before going into a more detailed development, says Ravindran. “The current extensions enable one to develop multi-agent learning algorithms for autonomous driving, so as to learn in an environment where the other drivers are also adapting. This, we believe, is crucial to develop more realistic and reactive driving behaviours.”

The basic requirements of autonomous driving like lane discipline and collision avoidance are met with relative ease today. “What stands between the current state-of-the-art and full-scale real-world adoption of the technology is the ability of the cars to negotiate complicated and unprecedented traffic situations with the precision of an expert human driver,” says Devashish Chakravarty, professor-in-charge of the Autonomous Ground Vehicle Research Group at IIT-Kharagpur. “Reinforcement learning in a multi-agent simulated environment has a promise to achieve just that and I think the work of this team is really the need of the hour.”

“We are pretty optimistic that MADRaS would facilitate solving some excellent research problems especially in the context of navigating traffic similar to Indian scenes and learning how to navigate in such traffic,” says Madhava Krishna, professor, and lab head at the Robotics Research Center, International  Institute of Information Technology (IIIT), Hyderabad.

Next stop: FAD

While partially autonomous cars are here, FAD (fully autonomous driving), with all the complicated scenarios that it will have to account for, is still a long way away – and could take decades to become a reality. The researchers envision inter-vehicle communication becoming ubiquitous and reliable with the advent of 5G data and phone services, enabling vehicles to transmit their intent to other neighbouring vehicles and develop situational awareness that’s more sophisticated than what humans are capable of.

As of now, autonomous driving is so hard that even time-tested technologies that go into autopilots for aircraft can’t solve it totally, says Naik. “Even the biggest players in this field like Tesla and Uber are facing fatal accidents, despite having the best talent. It’s an extremely hard problem to solve,” he says. In an accident that Tesla said was caused by the driver, a Joshua Brown died in May 2016 at Florida, US – the first such fatal accident involving an autonomous car.

Berkeley professor Michael Jordan recently published an essay in which he imagined what the self-driving car infrastructure of the future would look like. “The overall transportation system (an Intelligent Infrastructure system) will likely more closely resemble the current air-traffic control system than the current collection of loosely-coupled, forward-facing, inattentive human drivers,” he writes.

MADRaS enables machine learning and AI researchers to advance computer vision research algorithms, demonstrating their effectiveness in real-time decision-making through machine learning techniques, says Pradeep Dubey, Intel Fellow, and PCL-India director. “These algorithms will ultimately help navigate hazardous traffic scenarios to improve road safety in a way that is transformative,” he says.

For now, the researchers behind MADRaS are inviting the AI community to come and participate in developing the simulator further. The official blog lists a series of possibilities and problem statements related to multi-agent learning.

Some years from now, if scores of researchers want to play around with autonomous driving algorithms on an open source platform, they will have MADRaS as an option, and its creators to thank for.

---