New Delhi: Robots can learn how to move by watching themselves. They can then correct their body movements through observation. Teaching them this level of kinematic “self-awareness” was the subject of a new study published in Nature Machine Intelligence earlier this week.
Yuhang Hu, lead author of the study and a doctoral student at the Creative Machines lab at Columbia University, shared the study on LinkedIn and said, “We can’t baby our robots forever—they should learn on their own, like looking into a mirror to understand their bodies and motions.” The study, titled ‘Teaching robots to build simulations of themselves’ was published on 25 February.
Previous studies on robotics relied on virtual simulations to train the Artificial Intelligence (AI) model powering robots. These studies required high-quality simulators to enhance the adaptability of robots in a real-world environment. However, building such simulators is complex.
The team from Columbia University solved this by developing an autonomous robotic arm capable of creating its own simulation by observing its movement through a camera. This self-generated simulation is beneficial when robots have to continuously update their understanding of themselves as they experience damage over time.
In future, this capability could enable autonomous robots to adjust to damage and environmental changes on their own, reducing the need for constant human intervention, the study said.
“Our goal is a robot that understands its own body, adapts to damage, and learns new skills without constant human programming,” Yuhang said, according to Columbia Engineering news.
Using 2D images to create a 3D model
Under the new method, the robot creates a three-dimensional model of its movements using a two-dimensional (2D) camera.
The researchers have developed a new “self-supervised learning framework” which uses AI at three levels. The first level is the processing stage, where the raw video is captured by the robot. This raw video includes data on the robot’s joints and other areas of movement. It is converted into a binary image—a form which can be understood by the robot.
Finally, through predictive modelling, the robot—the robotic arm in this case—is instructed on how to move in the real world.
“Our method displayed consistent performance in both simulated and real-world environments. This proves the robustness of the approach, making it reliable in varied conditions,” the study said.
This development could pave the way for future autonomous robots that can perform self-repair and be used in advanced manufacturing systems.
For instance, if a robotic arm gets misaligned at a car factory, “instead of halting production, it could watch itself, tweak its movements and get back to welding—cutting downtime and costs,” said the researchers.
However, the dependence on two-dimensional images remains a significant limitation because it may produce inaccurate results.
“In future research, enhancing the camera resolution and integrating three-dimensional modelling sensors could provide a comprehensive evaluation and increase the accuracy in understanding of robot’s movements,” researchers said.
Sneha Yadav is an intern with ThePrint. She graduated from the first batch of ThePrint School of Journalism.
(Edited by Sanya Mathur)
Also Read: Surf’s up on Mars. Red planet once had ‘vacation-style’ beaches, suggest new findings