"In the movie 'Top Gun: Maverick,' Tom Cruise plays the role of Maverick, who is responsible for training young pilots to complete a seemingly impossible mission - flying their jets deep into rugged canyons, flying so low that they cannot be detected by radar, and then rapidly climbing out of the canyon at extreme angles to avoid crashing into the walls. Spoiler alert: With Maverick's help, these human pilots successfully complete their mission.
On the other hand, it is difficult for a machine to perform the same thrilling tasks. For example, for an autonomous aircraft, flying directly towards its target conflicts with the actions it needs to take to avoid colliding with canyon walls or staying undetected. Many existing artificial intelligence methods fail to overcome this conflict, known as the stability-obstacle problem, and cannot reach the target safely.
Researchers at the Massachusetts Institute of Technology (MIT) have developed a new technology that can better solve complex stability-obstacle problems compared to other methods. Their machine learning approach surpasses existing methods in terms of safety and improves stability tenfold.
In an experiment that would make Maverick proud, their technology effectively piloted a simulated jet through a narrow corridor without crashing into the ground.
"This is a long-standing, challenging problem. Many people are interested in it but don't know how to handle such high-dimensional and complex dynamics," said Chuchu Fan, an assistant professor of aeronautics and astronautics and a member of the Laboratory for Information and Decision Systems (LIDS) at MIT, and senior author of the paper.
Fan collaborated with Oswin So, the first author of the paper and a graduate student. The paper will be presented at the "Robotics: Science and Systems" conference in South Korea from July 10th to 14th.
Stability-Obstacle Challenge
Many methods solve complex stability-obstacle problems by simplifying the system, allowing them to use simple mathematics to solve it. However, the simplified results often cannot adapt to the dynamics of the real world.
A more effective technique is to use reinforcement learning, a machine learning method in which an agent learns through repeated trials and is rewarded for behaviors that bring it closer to the goal. But here, there are actually two goals - maintaining stability and avoiding obstacles - and finding the right balance is tedious.
The MIT researchers broke down the problem into two steps. First, they redefined the stability-obstacle problem as a constrained optimization problem. In this setup, solving the optimization allows the agent to reach and stabilize at its goal, meaning it stays within a certain region. By applying constraints, they ensure that the agent avoids obstacles, explained So.
"However, deep reinforcement learning is not designed to solve optimization problems in their extreme graph form, so we cannot directly apply it to our problem. We have to derive mathematical expressions that are applicable to our system. Once we have these new derivations, we combine them with some existing engineering tricks used by other methods," said So.
Race Against Time
To test their method, they designed control experiments with different initial conditions. For example, in some simulations, the autonomous agent needed to reach and stay within a target area while maintaining extreme maneuverability to avoid colliding with obstacles.
Compared to several baselines, their method was the only one able to stabilize all trajectories while ensuring safety. To further push their method, they used it to pilot a simulated jet, just like the scenes seen in the movie 'Top Gun.' The jet had to stabilize near the ground target while maintaining a very low altitude and staying within a narrow flight corridor.
This simulated jet model was open-sourced in 2018 and was designed as a test challenge by flight control experts. Could the researchers create a scenario where their controller couldn't fly? But the model was too complex to handle and still couldn't handle complex scenarios, said Fan.
The researchers' controller prevented the jet from crashing or stalling and stabilized to the target better than any benchmark method.
In the future, this technology could serve as a starting point for designing high-dynamic robot controllers that must meet safety and stability requirements, such as autonomous delivery drones. Or it could be implemented as part of a larger system. Perhaps the algorithm would only be activated when a car skids on a snowy road to help the driver safely return to a stable trajectory.
So added that the true highlight of their method is its ability to handle extreme situations that humans cannot handle.
"We should strive to provide safety and stability guarantees for reinforcement learning, which will give us confidence when deploying these controllers to critical mission systems. We believe this is the first step towards achieving this goal," he said.
Next, the researchers want to improve their technique to better consider uncertainty when solving optimization problems. They also want to study how the algorithm performs when deployed on hardware, as there may be mismatches between the model and the dynamics of the real world.
"Professor Fan's team has improved the reinforcement learning performance of safety-critical dynamical systems. The controllers they created not only achieve the goal but also ensure that the system can safely reach and stay at the goal indefinitely," said Stanley Bak, an assistant professor in the Department of Computer Science at Stony Brook University. "Their improved formulation allows for successful generation of safe controllers for complex scenarios, including a 17-state nonlinear jet aircraft model designed in part by researchers from the Air Force Research Laboratory (AFRL), which incorporates nonlinear differential equations with lift and drag tables."
Source: https://techxplore.com/news/2023-06-ai-based-approach-autonomous-robots.html"
Alphabet Q1 paid subscribers reached 350M, fueled by YouTube & Google One. YouTube ad revenue slightly missed at $9.88B amid Premium growth. Total Q1 revenue hit $109.9B, beating forecasts.
Anthropic releases Claude connectors for Adobe, Blender, and Ableton, enabling AI workflow automation and app control. It also sponsors the Blender Development Foundation.
Apple introduces an App Store model splitting 12-month plans into monthly installments. On iOS 26.4+, it lowers upfront costs without reducing total price, adding billing tracking.
Google launches Aadhaar digital IDs in Wallet for India, enabling secure QR/NFC verification. A phased partner rollout targets key sectors, while passport-based credentials expand globally.
YouTube tests "Ask YouTube," an AI tool blending text and video for complex queries. Initially for US Premium users, it supports interactive follow-ups, expanding Google's AI search.
Google grants US DoD classified AI access, joining OpenAI and xAI in defense deals amid Anthropic's dispute. Contractual safeguards against surveillance and lethal weapons remain unenforceable.
Mistral AI launches Workflows, a Temporal-powered orchestration platform for secure, Python-driven enterprise AI automation featuring human-in-the-loop approvals.