Shulou(Shulou.com)05/31 Report--

Before printing the manuscript, Microsoft researchers described a machine learning system that could infer the correct action directly from the camera image. Through simulation training, it learns to independently navigate real-world environments and conditions, including invisible situations, which makes it suitable for robots deployed in search and rescue missions. One day, it can help these robots identify people in need more quickly.

In a blog post released this week, the researchers wrote: "We hope to push existing technologies closer to human ability to interpret environmental clues, adapt to difficult environments and operate autonomously."We are interested in exploring what conditions are needed to build an autonomous system with a similar level of performance."

The team's framework clearly separates perceptual components (that is, getting what it sees from the control strategy (deciding what to do based on what it sees). Inspired by the human brain, it directly maps visual information to the correct control behavior, that is, by transforming a high-dimensional video frame sequence into a low-dimensional representation of the state of the world. The researchers say this two-phase approach makes the model easier to interpret and debug.

The team applied their framework to a small four-axle helicopter with a front-facing camera in an attempt to "teach" an artificial intelligence strategy so that it could navigate using only camera images during the game. They use a high-fidelity simulator called AirSim to simulate and train artificial intelligence and then deploy it to real drones without any modification. They use a framework called Cross-Modal variational Automated Encoder (CM-VAE) to generate a representation that can narrow the gap between simulation and reality.

The perception module of the system compresses the input image into the above low-dimensional representation, from 27648 variables to the most basic 10 variables. The decoded image provides a forward description that can be seen by the UAV, including the size and location of all possible doors and different background information.

The researchers tested the performance of the system on 45-meter-long S-shaped orbits and 40-meter-long circular orbits. They say that the strategy of using CM-VAE is significantly better than the end-to-end policy and AI that directly encode the next gate location. Despite the "strong" visual interference of background conditions, the UAV completed the course using a cross-modal perception module.

The two co-authors assert that the findings show "great potential" in practical applications. For example, the system can help autonomous search and rescue robots better identify humans, regardless of age, size, gender and race, so that robots have a better chance to identify and retrieve people in need.

"by dividing the perceptual-motor cycle into two modules and combining multiple data patterns into the perceptual training phase, we can avoid over-fitting our network to adapt to the irrelevant characteristics of the input data," the researchers wrote. "for example, in simulations and physical experiments, although the square doors are the same size, their width, color and even the intrinsic parameters of the camera do not exactly match."

The research comes after Microsoft released the UAV Challenge game, which allows quad artificial intelligence systems to compete in AirSim simulations. Microsoft introduced AirSim into its Unity game engine last year.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.