Microsoft is continuing to develop a machine learning system capable of making decisions based on visual data—similar to how humans reason through information and environmental cues autonomously. According to the company’s Autonomous systems division, these systems train deep neural nets using simulated data before testing them in real-world environments. These systems have been deployed in the DARPA Subterranean Challenge, which centers on using autonomous technology to assist first responders in rescue missions.
The researchers spearheading the project, called Team Explorer, have taken a two-stage approach to the system’s design. The components responsible for perception (making sense of what you see) and the control policy (deciding what to do based on what you see) are completely separated from each other. According to the team, this makes it easier to debug and interpret deep neural models.
The team has also explored drone racing research by studying first-person-view drone racing to better understand the system’s navigation abilities. Using the AirSim simulator during training, the researchers were able to simulate real-life environments and make adjustments according to these virtual representations. The key challenge was ensuring that the system would be able to clearly identify differences between the real world and the simulation.
A Cross-Modal Variational Auto Encoder (CM-VAE) was able to produce an environment nearest to real-life objects. Through this, the researchers were able to interpret the robot’s understanding of its environment through the cross-modal representations. According to the team, the system was able to outperform end-to-end control policies and networks without needing to reason through multipl data modalities.
Microsoft aims to further support research in autonomous systems that can become capable of solving real-world challenges. Team Explorer has already participated in the first two circuits of the DARPA Challenge, placing second at the February 2020 Urban Circuit and taking home first place in the September 2019 Tunnel Circuit. The Urban Circuit challenged robots to navigate courses that resembled complex urban underground infrastructures, such as stairs and other elevations, while the Tunnel Circuit involved underground tunnels that robots had to successfully navigate around while locating items.
According to the team, this technology can be beneficial in real-world applications, particularly in search and rescue settings. The robot can successfully recognize humans regardless of variations in age, size, ethnicity, gender, and other characteristics. Researchers will be exploring other modalities such as laser measurements and sound in learning environments. In addition, they will be using this study following the success of aerial testing to further enhance robotic abilities. The goal is to enable these robots to make quicker decisions, interpreting data in realtime while ensuring safety during the operation.
For more information, visit https://www.microsoft.com/en-us/research/blog/training-deep-control-policies-for-the-real-world/?OCID=msr_blog_controlreal_ai.
For more of the latest news and stories, check out how this soil sampling autonomous robot is boosting farm production here.