Cohort: 1 

Location: Seoul, South Korea 

Application Title: Ultra-efficient on-device AI for autonomous robotics 

Team Members: Lokwon Kim executive sponsor), Dong Jin Hyun (executive sponsor), Tim Park, Moon Sub Jin

The challenge 

Autonomous robots are increasingly used in indoor environments such as hospitals, offices and logistics facilities. These settings require robots to navigate safely around people, respond to dynamic obstacles and operate continuously on battery power. 

Early robotic systems relied on manual delivery or simple rule-based automation. These approaches reduced labour but struggled in complex environments where conditions change frequently. More advanced AI techniques improved perception and navigation, but they introduced a new constraint: computation. Running modern AI models typically requires graphics processing units (GPUs) or cloud connectivity, both of which consume significant energy and generate heat. 

For battery-powered robots, this creates a bottleneck. Power-hungry hardware shortens operating time, raises costs and limits deployment at scale. Reliance on cloud processing also introduces latency and connectivity risks, which are problematic for real-time decision-making in indoor spaces. 

The solution 

Custom AI hardware designed for edge robotics 

Hyundai partnered with DEEPX to develop an on-device AI solution that enables real-time inference without relying on cloud computing or energy-intensive GPUs. 

The solution centres on custom AI semiconductors designed specifically for inference tasks. These chips are optimised to run perception and navigation models efficiently within tight power and thermal constraints. Rather than adapting general-purpose hardware, the system is built to match the needs of mobile robotics from the outset. 

DEEPX provides the semiconductor architecture and optimisation software, while Hyundai integrates the technology into autonomous robotic platforms for real-world operation. 

How it works 

Running AI directly on the robot 

AI models used for tasks such as obstacle detection or path planning are optimised using proprietary software so they can run efficiently on the custom chips. This optimisation preserves accuracy while reducing computational load. 

A software development kit supports model conversion and deployment, allowing engineers to move from training to on-device execution without redesigning applications. Once deployed, inference runs entirely on the robot, removing dependence on external servers. 

Because computation happens locally, robots can respond immediately to their environment while maintaining low energy consumption. This makes continuous operation feasible in indoor settings where charging opportunities are limited. 

How it differs from conventional approaches 

Moving away from GPUs and cloud reliance 

Conventional robotic AI systems depend on GPUs or cloud processing to handle complex models. These approaches increase power draw and introduce delays, especially when connectivity is unreliable. 

The Hyundai–DEEPX solution replaces general-purpose hardware with purpose-built AI accelerators. By running models directly on the device, the system reduces energy use and avoids latency associated with remote processing. This makes advanced AI practical for small, mobile robots operating indoors. 

Impact 

• 240% performance increase compared with GPU-based systems 
• Energy consumption reduced to one-eighth of GPU requirements 
• Up to 90% reduction in total cost of ownership for deployed robots 



Why it matters 

Enabling practical autonomy at the edge 

As robotics expands beyond controlled environments, energy efficiency and real-time response become critical. The collaboration between Hyundai and DEEPX shows how AI hardware designed for edge use can support autonomous systems without the overhead of cloud infrastructure or high-power components. 

By lowering energy demands while maintaining performance, the approach supports wider deployment of autonomous robots in settings where reliability and cost control matter most.