Researchers from Shanghai-based Tongji University have unveiled an independently developed first-generation specialized computing chip for rolling optimization, known as the Moving Horizon Unit, which functions like an intelligent brain that can simultaneously observe, compute, and adjust when installed in autonomous vehicles, robots, drones, and other embodied intelligent devices.

According to the research team, the chip, developed using an innovative technological approach, will enable intelligent devices to efficiently perform end-to-end tasks, including real-time computation, decision-making, and command execution in complex environments. This advancement is poised to drive significant growth in the smart transportation and high-end intelligent equipment industries.

Chen Hong, a lead scientist on the team, explained that the concept of rolling optimization is an inherent human ability, involving continuous adjustments based on conditions and feedback to progressively achieve goals. In complex and ever-changing environments, autonomous vehicles, intelligent robots, and drones must constantly perceive changes, dynamically plan routes, and make real-time decisions, which are tasks that pose significant challenges for general-purpose chips.

"Our research aims to transform the human capability of adapting to environmental changes and uncertainties in a proactive way into a machine-computable hardware solution. By embedding this rolling optimization mechanism into a specialized chip, we enable vehicles to drive autonomously and robots to make independent decisions without relying on pre-set programs," said Chen, a professor from the School of Electronic and Information Engineering, Tongji University.

Leveraging three decades of experience in algorithm hardware integration, the team developed a specialized computing architecture for rolling optimization, over which they hold completely independent intellectual property rights.

Chen said the architecture incorporates several innovations at the foundational level, including precise computation by embedding predictive and optimization capabilities for the physical world into the hardware, enabling devices to dynamically plan for the future. Additionally, it learns quickly by innovatively integrating AI algorithms with real-world physical laws, significantly reducing the machines' reliance on vast amounts of data.

"Also, we have already developed an onboard controller, and conducted real vehicle integration tests," he said. Researchers said that the chip's core module offers both specialized customization and general integration capabilities, allowing it to be deeply adapted for various applications and embedded across multiple chip systems.