Researchers from the Hong Kong Polytechnic University have developed an agentic robotics system for the development of perovskite solar cells (PSCs). 

This system has been designed to enable a fully autonomous, closed-loop research framework.

Supported by a seven-layer AI architecture, it reads existing scientific literature, generates new chemical formulas, and translates them into machine code.

The demonstration showcased that their autonomous system fabricated perovskite solar cells with a power conversion efficiency of 27.0 percent. 

“Using this system, the team carried out 50,764 perovskite solar cell device experiments, achieved a champion power conversion efficiency of 27.0%, with a certified value of 26.5%,” the team noted.

A factory in a box

Despite its immense potential, perovskite solar cells remain hindered by a “trial-and-error” development process that is both labor-intensive and difficult to replicate. 

With over 100,000 experimental recipes tested, engineers still struggle to master complex formulas and volatile crystallization cycles that are hyper-sensitive to environmental changes. 

Current robots can gather data quickly, but aren’t smart enough to understand the science behind the results or make immediate adjustments to improve the process.

The new system integrates an AI architecture with physical robotics, including learning, generation, RecipeQA, fine-tuning, reasoning, evaluation, and optimization.

Moreover, a domain-specific Recipe Language Model (RLM) processes literature and experimental data to recommend, reason through, and refine chemical formulas.

The system features a network of 11 interconnected robotic boxes integrated into a single, cohesive framework.

Reportedly, its hardware is strategically partitioned: the first three robotic boxes manage chemical storage and precise material dispensing, while the remaining eight handle the core fabrication tasks, including spin-coating, laser processing, and deposition.

Equipped with an array of cameras and sensors, these units provide in situ characterization, capturing real-time data that is fed directly back into the AI loop to drive the model’s continuous evolution.

The 11 interconnected robotic boxes have been designed to control over 4,300 parameters.

It also uses a seven-layer AI architecture to continuously learn from both the scientific literature and its own experimental data, establishing a closed-loop workflow for synthesis, characterization, and optimization. 

As per the study, the power of this approach was demonstrated through over 50,000 experiments and the generation of 578 million data tokens. It ultimately achieved a power conversion efficiency of 27.0 percent.

101 functional modules

A real-time software-hardware bridge translates AI-generated recipes into machine instructions while monitoring the state of the 101 functional modules.

Through converting complex formulas into machine-readable commands and processing experimental results into structured feedback, it maintains a cycle of recommendation, robotic execution, and validation.

Moreover, this hardware evolution transforms existing, fragmented glovebox work into a unified, full-device fabrication system managed by a digital twin interface. 

This research provides a scalable foundation for materials intelligence by shifting away from fragmented, manual glovebox operations.

Ultimately, it could pave the way for autonomous, intelligent manufacturing in extreme or remote environments where human presence is impractical.

The findings were reported in the journal Engineering.