Bosung Kim | Searcharxiv

arXiv2026

How to Instruct Your Robot: Dense Language Annotations Power Robot Policy Learning

Scaling robot policy learning is bottlenecked by the cost of collecting demonstrations, while language annotations for existing demonstrations are comparatively cheap. We study language density as a lever for extracting more signal from a fixed robot or egocentric-video corpus. We introduce DeMiAn (Dense Multi-aspect Annotation), a two-stage approach that first re-labels demonstration segments with VLM-generated annotations along four complementary aspects: physical motion, scene composition, arm pose, and reasoning. A learned instructor then maps a task description and initial scene snapshot to a task-appropriate annotation at deployment, running asynchronously so generation latency is hidden behind policy execution. Across over 1M robot manipulation clips and 50K EgoVerse human-egocentric videos, DeMiAn improves both a vision-language-action policy and a video-based world-action model without collecting new demonstrations. On RoboCasa, the instructor raises success by 5 points over a task-only baseline and comes within 3 points of a per-task oracle. No fixed annotation aspect dominates across tasks, showing that selecting the right dense language matters. DeMiAn also improves composite-task and out-of-distribution performance, and shifts the compute-performance frontier in both mid-training and post-training after accounting for annotation-generation FLOPs. These results position dense re-annotation as a practical scaling lever for robot policy learning.

cs.RO↗

arXiv2025

Beyond Needle(s) in the Embodied Haystack: Environment, Architecture, and Training Considerations for Long Context Reasoning

We introduce $\infty$-THOR, a new framework for long-horizon embodied tasks that advances long-context understanding in embodied AI. $\infty$-THOR provides: (1) a generation framework for synthesizing scalable, reproducible, and unlimited long-horizon trajectories; (2) a novel embodied QA task, Needle(s) in the Embodied Haystack, where multiple scattered clues across extended trajectories test agents' long-context reasoning ability; and (3) a long-horizon dataset and benchmark suite featuring complex tasks that span hundreds of environment steps, each paired with ground-truth action sequences. To enable this capability, we explore architectural adaptations, including interleaved Goal-State-Action modeling, context extension techniques, and Context Parallelism, to equip LLM-based agents for extreme long-context reasoning and interaction. Experimental results and analyses highlight the challenges posed by our benchmark and provide insights into training strategies and model behaviors under long-horizon conditions. Our work provides a foundation for the next generation of embodied AI systems capable of robust, long-term reasoning and planning.

cs.AI↗

arXiv2025

On-device Sora: Enabling Training-Free Diffusion-based Text-to-Video Generation for Mobile Devices

We present On-device Sora, the first model training-free solution for diffusion-based on-device text-to-video generation that operates efficiently on smartphone-grade devices. To address the challenges of diffusion-based text-to-video generation on computation- and memory-limited mobile devices, the proposed On-device Sora applies three novel techniques to pre-trained video generative models. First, Linear Proportional Leap (LPL) reduces the excessive denoising steps required in video diffusion through an efficient leap-based approach. Second, Temporal Dimension Token Merging (TDTM) minimizes intensive token-processing computation in attention layers by merging consecutive tokens along the temporal dimension. Third, Concurrent Inference with Dynamic Loading (CI-DL) dynamically partitions large models into smaller blocks and loads them into memory for concurrent model inference, effectively addressing the challenges of limited device memory. We implement On-device Sora on the iPhone 15 Pro, and the experimental evaluations show that it is capable of generating high-quality videos on the device, comparable to those produced by high-end GPUs. These results show that On-device Sora enables efficient and high-quality video generation on resource-constrained mobile devices. We envision the proposed On-device Sora as a significant first step toward democratizing state-of-the-art generative technologies, enabling video generation on commodity mobile and embedded devices without resource-intensive re-training for model optimization (compression). The code implementation is available at a GitHub repository(https://github.com/eai-lab/On-device-Sora).

cs.CV↗

arXiv2025

On-device Sora: Enabling Training-Free Diffusion-based Text-to-Video Generation for Mobile Devices

We present On-device Sora, the first model training-free solution for diffusion-based on-device text-to-video generation that operates efficiently on smartphone-grade devices. To address the challenges of diffusion-based text-to-video generation on computation- and memory-limited mobile devices, the proposed On-device Sora applies three novel techniques to pre-trained video generative models. First, Linear Proportional Leap (LPL) reduces the excessive denoising steps required in video diffusion through an efficient leap-based approach. Second, Temporal Dimension Token Merging (TDTM) minimizes intensive token-processing computation in attention layers by merging consecutive tokens along the temporal dimension. Third, Concurrent Inference with Dynamic Loading (CI-DL) dynamically partitions large models into smaller blocks and loads them into memory for concurrent model inference, effectively addressing the challenges of limited device memory. We implement On-device Sora on the iPhone 15 Pro, and the experimental evaluations show that it is capable of generating high-quality videos on the device, comparable to those produced by high-end GPUs. These results show that On-device Sora enables efficient and high-quality video generation on resource-constrained mobile devices. We envision the proposed On-device Sora as a significant first step toward democratizing state-of-the-art generative technologies, enabling video generation on commodity mobile and embedded devices without resource-intensive re-training for model optimization (compression). The code implementation is available at a GitHub repository(https://github.com/eai-lab/On-device-Sora).

cs.CV↗

arXiv2024

On-off switchable nonreciprocal negative refraction in non-Hermitian photon-magnon hybrid systems

Photon-magnon coupling, where electromagnetic waves interact with spin waves, and negative refraction, which bends the direction of electromagnetic waves unnaturally, constitute critical foundations and advancements in the realms of optics, spintronics, and quantum information technology. Here, we explore a magnetic-field-controlled, on-off switchable, nonreciprocal negative refraction within a non-Hermitian photon-magnon hybrid system. By integrating an yttrium iron garnet film with an inverted split-ring resonator, we discover pronounced negative refraction driven by the system's non-Hermitian properties. This phenomenon exhibits unique nonreciprocal behavior dependent on the signal's propagation direction. Our analytical model sheds light on the crucial interplay between coherent and dissipative coupling, significantly altering permittivity and permeability's imaginary components, crucial for negative refraction's emergence. This work pioneers new avenues for employing negative refraction in photon-magnon hybrid systems, signaling substantial advancements in quantum hybrid systems.

quant-ph↗

arXiv2023

Zero-shot Triplet Extraction by Template Infilling

The task of triplet extraction aims to extract pairs of entities and their corresponding relations from unstructured text. Most existing methods train an extraction model on training data involving specific target relations, and are incapable of extracting new relations that were not observed at training time. Generalizing the model to unseen relations typically requires fine-tuning on synthetic training data which is often noisy and unreliable. We show that by reducing triplet extraction to a template infilling task over a pre-trained language model (LM), we can equip the extraction model with zero-shot learning capabilities and eliminate the need for additional training data. We propose a novel framework, ZETT (ZEro-shot Triplet extraction by Template infilling), that aligns the task objective to the pre-training objective of generative transformers to generalize to unseen relations. Experiments on FewRel and Wiki-ZSL datasets demonstrate that ZETT shows consistent and stable performance, outperforming previous state-of-the-art methods, even when using automatically generated templates. https://github.com/megagonlabs/zett/

cs.CL↗

arXiv2023

A Note on Location Parameter Estimation using the Weighted Hodges-Lehmann Estimator

Robust design is one of the main tools employed by engineers for the facilitation of the design of high-quality processes. However, most real-world processes invariably contend with external uncontrollable factors, often denoted as outliers or contaminated data, which exert a substantial distorting effect upon the computed sample mean. In pursuit of mitigating the inherent bias entailed by outliers within the dataset, the concept of weight adjustment emerges as a prudent recourse, to make the sample more representative of the statistical population. In this sense, the intricate challenge lies in the judicious application of these diverse weights toward the estimation of an alternative to the robust location estimator. Different from the previous studies, this study proposes two categories of new weighted Hodges-Lehmann (WHL) estimators that incorporate weight factors in the location parameter estimation. To evaluate their robust performances in estimating the location parameter, this study constructs a set of comprehensive simulations to compare various location estimators including mean, weighted mean, weighted median, Hodges-Lehmann estimator, and the proposed WHL estimators. The findings unequivocally manifest that the proposed WHL estimators clearly outperform the traditional methods in terms of their breakdown points, biases, and relative efficiencies.

stat.ME↗

arXiv2023

An Autonomous System for Head-to-Head Race: Design, Implementation and Analysis; Team KAIST at the Indy Autonomous Challenge

While the majority of autonomous driving research has concentrated on everyday driving scenarios, further safety and performance improvements of autonomous vehicles require a focus on extreme driving conditions. In this context, autonomous racing is a new area of research that has been attracting considerable interest recently. Due to the fact that a vehicle is driven by its perception, planning, and control limits during racing, numerous research and development issues arise. This paper provides a comprehensive overview of the autonomous racing system built by team KAIST for the Indy Autonomous Challenge (IAC). Our autonomy stack consists primarily of a multi-modal perception module, a high-speed overtaking planner, a resilient control stack, and a system status manager. We present the details of all components of our autonomy solution, including algorithms, implementation, and unit test results. In addition, this paper outlines the design principles and the results of a systematical analysis. Even though our design principles are derived from the unique application domain of autonomous racing, they can also be applied to a variety of safety-critical, high-cost-of-failure robotics applications. The proposed system was integrated into a full-scale autonomous race car (Dallara AV-21) and field-tested extensively. As a result, team KAIST was one of three teams who qualified and participated in the official IAC race events without any accidents. Our proposed autonomous system successfully completed all missions, including overtaking at speeds of around $220 km/h$ in the IAC@CES2022, the world's first autonomous 1:1 head-to-head race.

cs.RO↗

arXiv2022

Coexistence of coupling-induced transparency and absorption of transmission signals in magnon-mediated photon-photon coupling

Coexistence of coupling-induced transparency (CIT) and absorption (CIA) of signals in magnon-mediated photon-photon coupling was experimentally determined in a planar hybrid structure consisting of a yttrium iron garnet (YIG) film and three concentric inverted-split-ring resonators (ISRRs). The experimental observation of simultaneous CIT and CIA phenomena was ascribed to magnon-mediated photon-photon coupling between the individually decoupled ISRRs. In order to capture the generic physics of the observed interactions, we constructed an appropriate analytical model based on the balance between the coherent and dissipative multiple-paths interactions, which model precisely reproduced both the CIT and CIA experimentally observed from a single hybrid system. This work, promisingly, can provide guidance for design of efficient, flexible, and well-controllable photon-magnonic devices that are highly in demand for applications to quantum technologies currently under development.

quant-ph↗

arXiv2019

Abnormal anti-crossing effect in photon-magnon coupling

We report the experimental demonstration of an abnormal, opposite anti-crossing effect in a photon-magnon-coupled system that consists of an Yttrium Iron Garnet film and an inverted pattern of split-ring resonator structure (noted as ISRR) in a planar geometry. It is found that the normal shape of anti-crossing dispersion typically observed in photon-magnon coupling is changed to its opposite anti-crossing shape just by changing the position/orientation of the ISRR's split gap with respect to the microstrip line axis along which ac microwave currents are applied. Characteristic features of the opposite anti-crossing dispersion and its linewidth evolution are analyzed with the help of analytical derivations based on electromagnetic interactions. The observed opposite anti-crossing dispersion is ascribed to the compensation of both intrinsic damping and coupling-induced damping in the magnon modes. This compensation is achievable by controlling the relative strength and phase of oscillating magnetic fields generated from the ISRR's split gap and the microstrip feeding line. The position/orientation of an ISRR's split gap provides a robust means of controlling the dispersion shape of anti-crossing and its damping in a photon-magnon coupling, thereby offering more opportunity for advanced designs of microwave devices.

cond-mat.mes-hall↗

arXiv2017

Coupled breathing modes in one-dimensional Skyrmion lattices

We explored strong coupling of dynamic breathing modes in one-dimensional (1D) skyrmion lattices periodically arranged in thin-film nanostrips. The coupled breathing modes exhibit characteristic concave-down dispersions that represent the in-phase high-energy mode at zero wavenumber (k=0) and the anti-phase low-energy mode at the Brillouin zone boundary (k=kBZ). The band width of the allowed modes increases with decreasing inter-distance between nearest-neighboring skyrmions. Furthermore, the collective breathing modes propagate very well through the thin-film nanostrips, as fast as 200 ~ 700 m/s, which propagation is controllable by the strength of magnetic fields applied perpendicularly to the film plane. The breathing modes in 1D skyrmion lattices potentially formed in such nanostrips possibly can be used as information carriers in information processing devices.

cond-mat.mes-hall↗

arXiv2016

Coupled gyration modes and propagation in one-dimensional periodic skyrmion arrays as reliable information carrier

We report on a micromagnetic numerical simulation study of dynamic couplings between neighboring skyrmions in narrow-width nanostrips. We explored the coupled gyration modes and their characteristic dispersions in terms of the interdistance between the neighboring skyrmions. The application of perpendicular magnetic fields, importantly, allows for control/modification of the dispersion of the coupled modes. Coupled skyrmion gyration modes provide a new type of efficient, reliable, fast, low-power-consumption information-signal propagation in narrow-width straight and curved nanostrips, as driven predominantly by the exchange interaction between individual nano-scale skyrmions.

cond-mat.mes-hall↗