Alex Gyobeom Koo

the model stack.

foundation model. — how the base gets trained

imitation — solves: "how does a model learn to drive from nothing but human demonstrations?"

vlm transfer — solves: "how does a driving model get common sense it never saw in driving data?"

future-prediction — solves: "how do you learn from raw driving video with no labels, by predicting what happens next?"

self-play rl — solves: "how do you get past human level when there are no demos to copy?"

backbone. — what sees inside it

domain fine-tuning. — how it specializes

reasoning layer. — how it thinks

action head. — how it acts

compute strategy.

train big, distill small.

constraint-first.

odometry & transforms.

ROS2 transform trees and odometry.

coordinate frames and transforms for ROS-based mobile robots.

on-device inference.

solves: "how do you actually get a tensor through a compiled TensorRT engine on embedded hardware?"

solves: "why is CoreInference structured as preprocess/postprocess lifecycle stages instead of a bare function call?"

solves: "why doesn't TensorRT/ONNX Runtime give you preprocessing for free, and why do you have to build it yourself?"

model eval.

using forward_hooks to extract intermediate layer outputs from a pre-trained ResNet model in PyTorch.

record & replay.

creating a robot sensor data recording and replay system.

replay feature.

solves: "which time-aligned sensor inputs did the system associate with this inference tick?"

solves: "what temporal state or memory did the model carry into this inference tick?"

topic demuxer.

hermetic SIL test harness — solves: "you never fake the thing you're actually testing, everything around it."

ci/cd.

solves: "how do you keep CI fast and correct at monorepo scale, without running every test on every commit or trusting a hand-rolled 'which tests are affected' guess?"

Inevitable

What's In My Mind