What if, instead of hard-coding road rules into self-driving algorithms, AI agents were free to come up with their own ways of safely and efficiently sharing the road? That’s the premise of an international research team’s new paper, which looks at what happens when AI agents in driving environments are simply tasked with getting to destinations as quickly as possible without crashing into one another.
To ensure safe and efficient traffic flow, self-driving and human-driven vehicles alike are expected to abide by a set of road rules. These do not vary much globally: there are rigid requirements, for example that vehicles drive on a certain side of the road and stop at red lights; as well as more subtle social road rules like the implicit designation of faster lanes on highways. The researchers discovered that agents not provided with this information still tended to develop a set of similar road rules, especially when their perception of the road and other vehicles was imperfect, i.e. more humanlike.
“We use simple reward functions for multi-agent autonomous driving and present strong empirical evidence demonstrating that environment parameters like spatial agent density and perception noise incentivize the emergence of social road rules,” explains the paper’s first author Avik Pal, a computer science major student at the IIT Kanpur in India and visiting researcher at the University of Toronto and Vector Institute. His co-authors are affiliated with Nvidia, University of Toronto, and the Vector Institute.
Unlike prior work which focused on building driving simulators with realistic sensors that mimic LiDARs and cameras, the researchers explored whether and to what extent high-level design choices for a simulator — such as the definition of reward and perception noise — can determine if the agents trained there exhibit realistic behaviours.
The researchers equipped their agents with variably sparse LiDAR sensors, and rewarded them when they reached their target destination as quickly as possible without colliding with other agents in the scene. Their observations show that as the sensor signals of the agents trained in this multi-agent driving environment became more noisy, the agents begin to rely on constructs such as lanes, traffic lights, and safety distance, and thus were more likely to learn and observe road rules that mimic those commonly found in human driving systems.
Along with perception noise, the spatial density of agents was another important factor in the emergence of these road rules in the driving environment. The researchers thus propose dense multi-agent interaction and perception noise as critical considerations for simulators that seek to instill humanlike road rules in self-driving agents.
The research team also identified the parameters that led to the emergence of specific road rules such as obeying traffic signals, lane following, fast lanes on a highway, right of way, communication signals, pedestrians on a crosswalk, and minimum safety distance. “We hope that the lessons in state space, action space, and reward design from this work will transfer to simulators in which the prototypes for perception and interaction used in this work are replaced with more sophisticated sensor simulation,” Pal tweeted.
The researchers have released their PyTorch-based driving simulator and an accompanying suite of 2D driving environments to stimulate interest within the ML community to solve fundamental self-driving problems.
The paper Emergent Road Rules in Multi-Agent Driving Environments is on arXiv, and the code is on GitHub.
Reporter: Yuan Yuan | Editor: Michael Sarazen
This report offers a look at how China has leveraged artificial intelligence technologies in the battle against COVID-19. It is also available on Amazon Kindle. Along with this report, we also introduced a database covering additional 1428 artificial intelligence solutions from 12 pandemic scenarios.
Click here to find more reports from us.
We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.