crowd navigation compact ai training crowded environments a roundabout island disc with raised centre hub

Compact AI Training Method Reduces Navigation Conflicts in Crowded Environments

September 17, 2026
AI Research
Comment off

A paper accepted to ECCV 2026 by researchers at DGIST and KAIST identifies Skill Conflict, a phenomenon in which motion planning and motion prediction compete for the same weights inside the compact shared encoders that mobile robots run. Their fix, Disjoint Parameter Training with Sparse Merging, cut the JRDB collision rate from 0.0189 to 0.0091 on a 6.44 million parameter model. We read the paper, the supplementary tables and the ablations, including the three-way task split that made results worse.

One Material, Two Transistor Types: The Universal Charge Injector Pointing Toward Stacked AI Chips

September 16, 2026
AI Hardware
Comment off

A KAIST-led team has reported a single material — degenerately doped tin diselenide — that injects charge efficiently into both n-type and p-type atomically thin semiconductor channels, removing the need for a separate contact metal per polarity. The paper reports a drive current more than a thousand times higher than a nickel electrode in p-type tungsten diselenide, an on/off ratio above a billion in n-type molybdenum disulfide, a subthreshold swing below 70 mV per decade, and a monolayer CMOS inverter with a voltage gain near 340. This article explains the physics, converts the figures into engineering terms, places the result against earlier contact schemes, and sets out the distance between a laboratory inverter and a manufacturable chip.

AI Reduces Sensory Hallucinations, Even at Night or in Smoke: Inside KAIST’s DNA and MAD Methods

October 25, 2022
Artificial Intelligence
Comment off

KAIST has published two methods for cutting sensory hallucinations in multimodal AI: the failure where a model misreads what a sensor physically reports, or invents a perception in one channel because another channel suggested it. DNA optimisation teaches vision-language models the physics of thermal, depth and X-ray sensors using their own wrong answers as the training signal. MAD suppresses cross-modal interference at decoding time with no retraining at all. Here is what each method fixes, what the reported numbers do and do not establish, where sensory hallucinations cost the most in production, and what this line of work still leaves unsolved.

KAIST

Compact AI Training Method Reduces Navigation Conflicts in Crowded Environments

One Material, Two Transistor Types: The Universal Charge Injector Pointing Toward Stacked AI Chips

AI Reduces Sensory Hallucinations, Even at Night or in Smoke: Inside KAIST’s DNA and MAD Methods

Tags