AI training method helps robots carry lab-learned skills into real-world tasks

Robots are trained for specific tasks, such as cutting, using simulation. However, collecting real-world data is expensive, slow, and sometimes unsafe, particularly for tasks involving physical interaction. A new AI-based method co-developed by Aston University's Dr. Alireza Rastegarpanah could revolutionize the way advanced robotic systems are trained for real-life tasks, making them more practical and reliable.

Dr. Rastegarpanah, assistant professor in applied AI and robotics at Aston, co-led research with Jamie Hathaway from the University of Birmingham's Extreme Robotics Lab to overcome the "sim-to-real gap." This is a longstanding challenge in robotics, referring to the difference between how robots behave in simulation and how they behave in the real world, where there is variability, for example in materials, forces or sensor noise. This leads to unreliability.

The goal of the research, published in Scientific Reports, was to develop a method that combined the efficiency of simulation with the realism of physical environments, enabling robots to adapt without requiring large amounts of additional data.

By using AI to generate variations in conditions, the new training technique allows robots to transfer skills learned in simulation into the real world much more reliably, using only a small amount of real-world data. A robot can learn a complex task in a virtual environment, such as cutting or manipulating materials, and then adapt that knowledge to work effectively in real-world conditions, even when those conditions are uncertain or previously unseen.

Dr. Rastegarpanah says that the method demonstrates that it is possible to achieve stable, efficient, and adaptive robot behavior without requiring extensive real-world training. It could significantly reduce development time, cost, and risk. The impact is particularly strong in areas where robots must operate under uncertainty. This includes recycling and circular economy systems, such as battery disassembly, advanced and flexible manufacturing, and hazardous environments such as nuclear decommissioning.

Dr. Rastegarpanah said, "This work shows that we can move beyond purely simulation-based training and achieve reliable performance in real-world conditions with minimal additional data. Our long-term vision is to enable plug-and-play intelligent robotic systems that can be trained in simulation and rapidly deployed in new environments with minimal reconfiguration. This could significantly accelerate innovation in areas such as sustainable manufacturing, recycling, and autonomous industrial systems."

Provided by Aston University