A brief self-introduction.

I'm Jaehong Oh — a robotics software engineer and research intern on the ROBOTIS Perception Team, studying Mechanical Engineering at Soongsil University in Seoul. My work sits at the intersection of hardware and a question that engineering alone doesn't close: not just what systems can do, but what they can know — about their state, their environment, the gap between a model and what it models. That question came to me before I wrote a line of code, through sustained attention in a very different medium, where the only verification is whether what you wrote actually says what you think. The question that kept surfacing: what persists when something cannot fully observe itself? I haven't answered it, but it has organized almost everything I've built since. Like the hollow inside a ring — not empty space, but the structure that makes the ring what it is.

Three projects mark the progression. TurtleBot4 was a seven-person industrial safety monitoring system — real-time detection, 4D state estimation tracking from position through jerk, MQTT-coordinated robot fleet — 92.3% accuracy at 350ms end-to-end. Soma Cube applied reinforcement learning to robotic assembly: a Masked DQN with legal-action masking, converging from a 54% baseline to 96.1% success over 105,300 episodes. The structural insight was that collision avoidance has to be built into the action space — you can't fine a system out of a configuration it was never prevented from entering. ONN is a topology-preserving network where the latent state carries explicit ontological structure, achieving 99.75% topology preservation across 3 million nodes and 14.7% perplexity improvement over Transformer baselines (arXiv 2506.19277). Each project asked the same question in a different language.

The theoretical work — Soft Cognitive Cohesion and Ontology Neural Networks — grew from the engineering rather than alongside it. SCC formalizes what exists before the world is parsed into separately identified things: a soft cohesion field over a relational support space, from which existence emerges as a gauge-invariant topological invariant — not assumed, but derived. ONN makes this learnable: the latent state carries explicit ontological structure, and delay-robust feedback closes the loop on top of that topology. Most models encode structure the way memory encodes a dream — you know something was there, but the shape is gone. These frameworks try to keep the shape visible. The work was accepted at the International Journal of Topology in 2026.

The direction I'm working toward runs on bidirectional arrows. Hardware failures are data about theory: an arm oscillating at a control boundary points to missing topology in the state space; a grasp failing in an unseen configuration reveals a gap in the relational geometry. When the theory sharpens, it generates specific hardware requirements in return — not tuning criteria, but structural specifications that change what you build. My goal is an architecture where ontological structure is explicit at every layer, from raw sensor input through learned latent representations to the control interface. The arrows run in both directions, and the most interesting problems live exactly at that crossing.