The Founder Behind Devol Robots, the Startup Rethinking How Robots Learn

Most work in manufacturing, logistics, and assembly still relies on human hands. Despite decades of investment in automation, tasks involving variation, tight tolerances, or unstructured environments remain resistant to robotics.

Sze Yuan Cheong witnessed this challenge firsthand. After spending more than a decade building and operating four manufacturing businesses, he became convinced that the gap between what robots could do and what factories actually needed was not merely an engineering issue—it was fundamentally misframed. That conviction led him to co-found Devol Robots, a physical AI startup developing a force-based world model for robotic manipulation.

Now approaching three years since incorporation, Devol Robots employs an 18-person team and is deploying its model with U.S.-based industrial clients.

An Early Interest In Entrepreneurship

Born in Singapore and raised in Malaysia, Cheong moved to the United States for college, where he quickly became interested in startups. During a summer trip home, he was pitched a business idea on his second day back. By the third day, he had launched the company. At 21, he chose not to return to complete his degree, instead running the venture full-time.

Over the following years, Cheong built four traditional manufacturing-related businesses. His academic background in industrial and process engineering provided deep insight into how goods are produced—and where production systems fail. While financially successful, he eventually found the work lacking long-term purpose.

That restlessness, combined with a lifelong fascination with technology, led him to reconsider what he wanted to build next.

Finding the Pain Points in Robotic Training

Years of operating manufacturing companies gave Cheong a clear understanding of the industry's persistent bottlenecks. Assembly lines, logistics handoffs, and tasks involving modest variation remain largely manual despite automation investments. The disconnect between robotics research and factory-floor deployment represented both a problem and an opportunity.

The turning point came when Cheong reconnected with Elijah Yi Herng Ong, a researcher who had completed a master's degree focused on reinforcement learning for robotic grasping and had been admitted to Stanford's PhD program, which he ultimately declined. Ong instead joined a robotics startup in Austin, Texas, where he worked on building a robotic arm from the ground up. There, he encountered impedance control—a method of guiding robots through stiffness and compliance, inspired by how humans regulate muscle tension.

Together, Cheong and Ong developed the mathematical foundations for describing physical embodiment through force-based control and founded Devol Robots to commercialise the research. Their aim: build a foundational model that teaches robots to interact physically with objects—not by watching videos, but by learning from forces, torques, and contact dynamics, much like humans learn through touch.

Devol Robots: A Different Approach to Automation

The prevailing method in robotic learning today is often described as the vision-language-action pipeline. In simplified terms, a robot's movements are tokenised, processed through a large language model, and used to predict future visual states, which are then translated back into motor commands. While functional, this method infers physical interaction primarily from vision.

Devol Robots takes a different approach. Its model integrates torque and force readings at each joint with stiffness parameters and visual data. Instead of reconstructing physics from pixels, the system learns directly from embodied physical interaction.

Training relies on a recurrent neural network with temporal sequencing rather than diffusion-based modeling. This allows the system to better capture acceleration, rate of change, and contact dynamics over time.

Vision remains important but serves primarily to interpret context and outcomes rather than simulate physics. Because control is abstracted into stiffness parameters, the model can generalise across different robotic bodies and tasks—reducing the need for retraining with each new embodiment.

Deploying at Industrial Scale

Devol Robots is currently deploying its technology with a U.S.-based optics manufacturer producing high-end lenses across thousands of variants. Each lens requires a unique jig and fixture with unpredictable six-degree-of-freedom (6D) orientations. The variability makes pre-programmed robotic placement impractical.

Standard vision-based models achieve roughly 50% reliability under such conditions—far below industrial requirements. Some competitors attempt to compensate through human teleoperators using VR interfaces, but the precision required for lens placement makes remote control inefficient and risky. Each failed attempt can damage high-value components.

Devol's physics-grounded model adapts in real time using force feedback. By detecting contact dynamics, the robot adjusts its strategy without human intervention.

The company's 18-person team—largely based in Malaysia—focuses on recruiting high-performing researchers drawn to frontier problems in robotics and embodied AI.

Changing How Robots Interact With the Physical World

Nearly three years after its founding, Sze Yuan Cheong and Devol Robots are betting that the future of industrial automation lies not in better cameras, but in teaching machines to feel what they are doing.

With active industrial deployments, a growing research team, and a model architecture designed to generalise across embodiments and tasks, the company aims to offer a physics-grounded alternative to prevailing robotic AI approaches—bringing robots closer to reliable, adaptable performance on real factory floors.