A new artificial intelligence-based robotics training method could significantly improve how industrial robots are prepared for real-world tasks such as cutting, assembly, and materials handling, according to new research.
The study, co-led by Dr Alireza Rastegarpanah, Assistant Professor in Applied AI and Robotics at Aston University, and Jamie Hathaway of the University of Birmingham’s Extreme Robotics Lab, introduces a sim-to-real transfer framework designed to address one of robotics’ most persistent barriers: the gap between how well robots perform in a lab environment and how they perform in the real world.
Robots are typically trained in simulation because real-world data collection is expensive, time-consuming, and sometimes unsafe, particularly in contact-rich tasks like cutting or disassembly.
However, policies trained purely in simulation often struggle when transferred to physical environments, limiting their industrial usefulness.
This mismatch – described by the researchers as the ‘Sim-to-real gap – tends to mean even the best made robots currently behave unreliably in real life settings due to a variability in materials, forces, and sensor noise.
Instead, the researchers used a new method which seeks to combine the efficiency of simulation with the realism of real-world data. It does this by using AI-based techniques to generate realistic variations in training data, allowing robots to adapt more effectively when deployed in unpredictable conditions.
They found that by integrating neural style transfer with variational autoencoders and reinforcement learning, the system transforms simulated robot trajectories into forms that more closely resemble real-world behaviour.
These “surrogate” real-world trajectories are then used to train control policies capable of handling physical uncertainty without requiring large-scale real-world datasets.
“This work shows that we can move beyond purely simulation-based training and achieve reliable performance in real-world conditions with minimal additional data,” Dr Rastegarpanah said. “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.”
According to the researchers, the new approach allows a robot to learn a complex task in a virtual environment, such as cutting or manipulating materials, and then transfer that knowledge to physical environments with minimal additional training. Crucially, this enables adaptation even when the robot encounters previously unseen conditions.
The research was supported by the REBELION project, funded by UK Research and Innovation (UKRI) as part of a wider European collaborative initiative focused on automated and safe lithium-ion battery recycling. This context is particularly relevant given growing industrial interest in robotics for circular economy applications, including battery disassembly and material recovery.
In experimental trials using a KUKA LBR iiwa robotic arm performing cutting tasks across multiple materials, the proposed method demonstrated improved stability, faster task completion times, and more consistent behaviour compared to several baseline approaches. These included un-adapted simulation-trained policies, CycleGAN-based translation methods, and conditional variational autoencoder models.
The researchers report that their method achieved strong performance while using only a small amount of real-world data, a key advantage in industrial settings where physical trials are costly or hazardous. Only 148 real-world trajectories were required for adaptation, compared with tens of thousands of simulated episodes used for initial training.
Beyond performance improvements, the study also found that the system produced smoother and more stable robot actions under challenging conditions such as uneven surfaces or path deviations. This is particularly important in industrial cutting and machining tasks, where unstable control can lead to tool wear, reduced precision, or equipment damage.
Despite its promising results, the researchers acknowledge limitations. The method does not explicitly enforce physical feasibility constraints during trajectory generation, meaning some outputs may not fully align with real robot kinematics. They also note that performance depends on the quality and diversity of available real-world data, particularly in the style transfer stage.
