Teahose.
SIGN IN
NEW HERE — WHAT TEAHOSE DOES
We read the entire AI & tech firehose — so you don't have to.
PODPodcastsAll-In, No Priors, Acquired…
NEWNewslettersStratechery, Newcomer…
PAPPapersPhysical AI research
PHProduct Huntdaily launches
VCInvestor ScoutSequoia, a16z, Benchmark…
CLAUDE DISTILLS →
7 reads, 30 sec each — free, 6 AM ET.
+ a live graph of the companies, people & themes underneath.
HOME/PEOPLE/BOSUNG KIM
// PERSON

Bosung Kim

ROLE RESEARCHERAT NVIDIAMENTIONS 5LAST SEEN MAY 29, 2026
// BIO

Bosung Kim is a PhD candidate in Computer Science and Engineering at the University of California, San Diego, currently serving as a research intern at NVIDIA Research's Learning and Creativity (LACR) team, where he has been based since February 2026. He is advised by Prithviraj Ammanabrolu at UC San Diego's PEARLS Lab. Kim is best known for his work on dense language annotation for robot policy learning, including the DeMiAn (Dense Multi-Aspect Annotation) framework, which improves vision-language-action model performance by enriching existing robot demonstration corpora with structured language annotations without requiring new demonstrations.

// RECENT MENTIONS
// SIGNALS
5 SIGNALS
01
product·arXiv Physical AI·MAY 29, 2026

DeMiAn: Dense Multi-Aspect Annotation for Robot Policy Learning

Source
02
mention·arXiv Physical AI·MAY 29, 2026

On RoboCasa, the best fixed annotation type (Physical Motion) raised success rate from 44% to 46%, and the learned instructor pushed that to 49% — within 3 points of a theoretical per-task oracle at 52%.

Source
03
mention·arXiv Physical AI·MAY 29, 2026

At 1M-clip scale on MolmoBot, the compute cost for one annotation pass is roughly $1,100 and 5×10¹⁹ FLOPs. Dense annotation then matches the unannotated baseline on two MolmoSpaces task families while using ~62% less training compute — saving ~1.3×10²⁰ FLOPs.

Source
04
mention·arXiv Physical AI·MAY 29, 2026

Async tracks sync within fractional points on SR (49.0% vs 49.5%) while injecting the instruction into a rollout already in progress.

Source
05
mention·arXiv Physical AI·MAY 29, 2026

DeMiAn VLA: We use the open-source openpi 0.5 as the VLA backbone, a PaliGemma vision-language backbone paired with a flow-matching action expert.

Source

AI-extracted from podcast / newsletter / paper summaries. May contain errors.