Robots can perceive the world, but turning perception into safe and efficient action in real time remains a difficult engineering problem, especially when everything must happen directly on-device.
That is the challenge addressed by Simfero, the project developed by Iuliia Gorshkova within the Second Open Call of dAIEDGE.
Rather than relying on cloud-based processing, Simfero focuses on bringing intelligence closer to the robot itself, positioning AI between sensor input and action output. The goal is to enable more adaptive and safer decision-making under strict constraints of power, memory and latency.
A key aspect of the work is understanding how AI behaves when deployed on real hardware. Instead of assuming ideal performance, the project measures execution directly on edge platforms—analyzing inference speed, energy consumption and bottlenecks in real conditions. This hardware-aware perspective becomes essential when optimizing for embedded systems such as FPGA-based architectures.
In this context, experimental results show promising performance, reaching very low inference latency and minimal energy usage, suggesting that FPGA-based deployment can be a viable path for real-time robotic intelligence where efficiency is critical.
During the programme, Iuliia collaborated with German Research Center for Artificial Intelligence, whose support and expertise helped structure the project around realistic hardware constraints and evaluation methodologies. We would like to sincerely thank DFKI for their role as hosting institution.
A defining element of the work was access to the dAIEDGE-VLab environment, which allowed the team to move beyond simulation and directly evaluate performance under realistic deployment conditions. This practical validation was key in understanding the trade-offs between speed, power and accuracy.
Simfero reflects the core vision of dAIEDGE: enabling intelligent systems that are not only advanced in design, but also efficient, measurable and deployable in the real world.