Leading the AI Airspace: MIT’s Innovative Method Ensuring Safe and Dependable Autopilots for Flight

78

MIT’s scientists have created an AI-centric technique that enhances safety and balance in self-governing robots, effectively solving the ‘stabilize-avoid’ dilemma. The two-fold method integrates deep reinforcement learning with mathematical optimization and has proven successful on a simulated jet aircraft, indicating potential uses in dynamic robots requiring safety and stability, like autonomous delivery drones.

A freshly designed AI-based strategy for controlling autonomous robots meets the often opposing goals of safety and balance.

In the movie “Top Gun: Maverick,” Tom Cruise’s character, Maverick, trains novice pilots to carry out a seemingly impossible task – fly their jets deep into a rocky canyon, maintaining such a low altitude that they escape radar detection, then swiftly ascend at an extreme angle, evading the canyon walls. With Maverick’s guidance, the human pilots manage to achieve their mission.

However, a machine might find this adrenaline-fueled task challenging. To an autonomous aircraft, the simplest route to the target conflicts with its necessary actions to prevent collision with the canyon walls or remain undetected. Current AI methods often fail to resolve this ‘stabilize-avoid’ problem and cannot reach their goal safely.

MIT’s team has devised a machine learning technique that can navigate a car or fly a plane autonomously through a challenging ‘stabilize-avoid’ situation, in which the vehicle must maintain its trajectory while dodging obstacles.

The researchers at MIT have created a new technique that can tackle complex stabilize-avoid problems more effectively than other strategies. This machine-learning approach matches or surpasses the safety of existing methods and improves stability tenfold, allowing the agent to reach its goal and remain stable within it.

In an experiment mimicking a scene from “Top Gun,” the researchers successfully piloted a simulated jet aircraft through a tight passage without any crash.

“This has been a tough, longstanding problem. Many have tried to solve it but were unable to manage the complex dynamics,” says Chuchu Fan, the Wilson Assistant Professor of Aeronautics and Astronautics, a member of the Laboratory for Information and Decision Systems (LIDS), and the paper’s senior author.

Fan is assisted by lead author Oswin So, a graduate student. The paper will be shared at the Robotics: Science and Systems conference.

The stabilize-avoid dilemma

Many methods try to solve intricate stabilize-avoid problems by simplifying the system, using basic math to solve it, but these simplified solutions often fail in real-world conditions.

More effective methods use reinforcement learning, where an agent learns through trial and error, rewarded for actions bringing it closer to a goal. But balancing between two objectives – stability and obstacle avoidance – can be tiresome.

MIT’s researchers divided the problem into two stages. They first reframed the stabilize-avoid problem as a constrained optimization problem. By adding constraints, they ensure the agent dodges obstacles while reaching and stabilizing at its goal.

In the second stage, they recast this constrained optimization problem into an epigraph form and solved it using a deep reinforcement learning algorithm, overcoming the difficulties other methods encounter when using reinforcement learning.

“But deep reinforcement learning isn’t equipped to solve the epigraph form of an optimization problem, so we had to create mathematical expressions suitable for our system. Once we obtained these new derivations, we combined them with some established engineering tricks used by other methods,” So says.

No room for runner-ups

In various control experiments designed to test their approach, their method was the only one capable of maintaining all trajectories safely. They further tested their method by simulating a jet flight similar to one seen in a “Top Gun” movie, where the jet had to stay low and within a narrow flight corridor while stabilizing towards a ground-level target.

The simulation model, which had been open-sourced in 2018 and designed as a test challenge by flight control experts, proved complicated to handle and still couldn’t manage complex scenarios, Fan says. However, MIT’s controller successfully stopped the jet from crashing or stalling and stabilized it far better than any of the baselines.

This method could form the basis for creating controllers for highly dynamic robots that need to meet safety and stability standards, like autonomous delivery drones. It might also be incorporated into a larger system, perhaps triggered only when a car skids on a snowy road to help navigate back safely.

“The aim is to provide reinforcement learning with the safety and stability assurances needed for deploying these controllers on mission-critical systems. We believe this is a promising initial step toward that goal,” So says.

Moving forward, the team plans to refine their technique to better consider uncertainty during optimization and assess how well the algorithm performs on hardware due to discrepancies between model dynamics and real-world dynamics.

“Professor Fan’s team has enhanced reinforcement learning performance for dynamic systems where safety is paramount. Instead of merely reaching a goal, they develop controllers that ensure the system can achieve its target safely and maintain it indefinitely,” says Stanley Bak, an assistant professor in the Department of Computer Science at Stony Brook University, who was not part of this research.

The research is partially funded by MIT Lincoln Laboratory under the Safety in Aerobatic Flight Regimes program.

Frequently Asked Questions (FAQs) about Autonomous Robots Safety

More about Autonomous Robots Safety

• MIT News
• Artificial Intelligence at MIT
• MIT Laboratory for Information and Decision Systems (LIDS)
• Reinforcement Learning
• Safety in Autonomous Systems
• Autonomous Drones