MIT scientists have developed a new visual-based artificial intelligence (AI) system that can teach themselves how to control virtually every robot without using sensors or pre-suppression.
The system uses cameras to collect data about the architecture of a particular robot. This is how humans learn about themselves as they move in almost the same way as they use their own eyes to learn about themselves.
This allows AI controllers to develop self-learning models for operating any robot. It essentially gives the machine a sense of physical self-awareness.
You might like it
Researchers achieved this breakthrough by creating a new control paradigm that uses cameras to map video streams in the “Visuomotor Jacobian field” of robots.
AI models can predict precise motion movement. This allows non-traditional robotics, such as soft robotics and robots designed with flexible materials, to be transformed into autonomous units in just a few hours of training.
“Think about how to control your fingers: you sway, you observe and adapt,” Sizhe Lester Li, a doctoral student at MIT CSAIL and a lead researcher for the project, explained in a press release. “That’s what our system does. It experiments with random actions and knowing which controls move which parts of the robot.”
Related: Scientists burn, plunge and slice new robotic skin that can “feel it all”
A typical robotics solution relies on precision engineering to create machines to precise specifications that can be controlled using pre-trained systems. These require expensive sensors and AI models developed with hundreds or thousands of hours of fine tuning to predict any possible permutation of motion. For example, gripping objects with attachments like handles remains a challenging challenge in both machine engineering and AI system control.
Understand the world around you
In contrast, the “Jacobian Field” mapping camera solution provides a low-cost, high-fidelity solution to the challenges of automating robotic systems.
The team published their findings in the Nature Journal on June 25th. In it, they said that the work is designed to mimic the way the human brain is learning to control machines.
The ability to learn and reconstruct 3D configurations and predict movement as a function of control is derived solely from vision. According to the paper, when controlling a robot with a video game controller, “people can choose and place objects within minutes,” and “the only sensor needed is our eyes.”
The system’s framework was developed using 12 to 3 hours of multi-view video.
This framework consists of two important components: The first is a deep learning model that allows robots to fundamentally determine where they reside in three-dimensional space. This allows you to predict how a particular move command will change its position when it is executed. The second is a machine learning program that converts general-purpose movement commands into code that a robot can understand and execute.
The team tested new training and control paradigms by benchmarking the effectiveness of traditional camera-based control methods. The Jacobian Field Solution outperformed existing 2D control systems with accuracy, especially when teams introduced visual occlusion that caused the old methods to enter into a failed state. However, the machine using the team’s method successfully created a navigable 3D map even if the scene was partially blocked by a random clutter.
When scientists developed the framework, it was applied to a variety of robots with different architectures. The final result was a control program that did not require further human intervention to train and operate the robot using only a single video camera.
Source link
