arXiv subjects

Daehee Lee

Publications and source records attributed to Daehee Lee.

13 recordsLinked to original sources

arXiv2026

Efficient Skill Grounding via Code Refactoring with Small Language Models

Effective skill grounding is essential for deploying reusable skills in embodied agents, as even minor embodiment or environmental differences can render an entire skill incompatible. This challenge is particularly pronounced in embodied settings, where agents must operate in dynamic, partially observable environments without access to large language models (LLMs). In this setting, reliance on LLMs is impractical, while small language models (sLMs) remain insufficient for the effective skill grounding required for reliable long-horizon control. We present RECENT, a refactoring-centric agent framework that enables efficient skill grounding with sLMs by decoupling skill semantics from embodiment- and environment-specific execution binding. By representing skills as executable code, RECENT preserves the semantic intent encoded in a skill's control structure while grounding it by modifying only execution bindings through localized refactoring, rather than regenerating code from scratch. We evaluate RECENT across diverse skill grounding scenarios spanning multiple robot embodiments in dynamic environments, demonstrating robust long-horizon performance when deployed with an sLM. Across all scenarios, RECENT achieves the best performance among sLM-based Code-as-Policies (CaP) methods and matches the task performance of LLM-based CaP.

cs.AI↗

arXiv2026

A Recipe for Stable Offline Multi-agent Reinforcement Learning

Despite remarkable achievements in single-agent offline reinforcement learning (RL), multi-agent RL (MARL) has struggled to adopt this paradigm, largely persisting with on-policy training and self-play from scratch. One reason for this gap comes from the instability of non-linear value decomposition, leading prior works to avoid complex mixing networks in favor of linear value decomposition (e.g., VDN) with value regularization used in single-agent setups. In this work, we analyze the source of instability in non-linear value decomposition within the offline MARL setting. Our observations confirm that they induce value-scale amplification and unstable optimization. To alleviate this, we propose a simple technique, scale-invariant value normalization (SVN), that stabilizes actor-critic training without altering the Bellman fixed point. Empirically, we examine the interaction among key components of offline MARL (e.g., value decomposition, value learning, and policy extraction) and derive a practical recipe that unlocks its full potential.

cs.LG↗

arXiv2025

Investigating the Timing Behavior of Compton Scattering in BGO for Time-of-Flight PET

Bismuth germanate (BGO) is gaining renewed attention as a viable material for hybrid Cherenkov/scintillation time-of-flight positron emission tomography (TOF-PET) detectors. While single-crystal studies have demonstrated excellent timing resolution by leveraging prompt Cherenkov photons, practical detector modules based on pixelated arrays introduce a high prevalence of inter-crystal scattering (InterCS) events, complicating timing accuracy. In this study, we experimentally investigated the impact of InterCS on BGO Cherenkov timing using a dual-pixel detector coupled to a segmented SiPM readout. Events were classified into full-energy deposition (FED; primary crystal 511 keV absorption), InterCS, and penetration types via energy-weighted positioning and validated using GATE simulations, which also revealed that over 25% of the experimentally identified full-energy events involved intra-crystal scatter (IntraCS). For InterCS events, the optimal timestamp selection was achieved by choosing the earlier of the two timestamps, yielding a coincidence timing resolution of 221 ps FWHM (831 ps FWTM)-markedly worse than the 184 ps (603 ps FWTM) obtained for FED events. Furthermore, prompt photon yield was found to decrease measurably due to energy splitting: InterCS events averaged 4.73 detected photons in the first 1 ns, compared to 5.76 for FED events. These results emphasize the importance of incorporating time-aware, per-pixel timestamping strategies in pixelated BGO TOF-PET systems to maintain optimal timing performance in the presence of scatter.

physics.ins-det↗

arXiv2025

Multi-agent Coordination via Flow Matching

This work presents MAC-Flow, a simple yet expressive framework for multi-agent coordination. We argue that requirements of effective coordination are twofold: (i) a rich representation of the diverse joint behaviors present in offline data and (ii) the ability to act efficiently in real time. However, prior approaches often sacrifice one for the other, i.e., denoising diffusion-based solutions capture complex coordination but are computationally slow, while Gaussian policy-based solutions are fast but brittle in handling multi-agent interaction. MAC-Flow addresses this trade-off by first learning a flow-based representation of joint behaviors, and then distilling it into decentralized one-step policies that preserve coordination while enabling fast execution. Across four different benchmarks, including $12$ environments and $34$ datasets, MAC-Flow alleviates the trade-off between performance and computational cost, specifically achieving about $\boldsymbol{\times14.5}$ faster inference compared to diffusion-based MARL methods, while maintaining good performance. At the same time, its inference speed is similar to that of prior Gaussian policy-based offline multi-agent reinforcement learning (MARL) methods.

cs.LG↗

arXiv2025

Scintillator-integrated microchannel plate photomultiplier tubes for ultrafast timing over keV-GeV energy scales

Precise measurement of radiation has long played a vital role in a wide range of research and industrial fields, from fundamental physics beyond the Standard Model to medical imaging such as time-of-flight positron emission tomography. Developing radiation detectors that achieve high timing precision-on the order of a few tens of picoseconds-and energy measurement capabilities remains indispensable yet challenging. In this study, we developed two types of scintillator-integrated microchannel plate photomultiplier tubes (SCI-IMPs), one incorporating barium fluoride, and the other bismuth germanate, to enable simultaneous high-precision timing and energy measurements. To evaluate their performance over a wide energy range from keV- to GeV-scale, electron-positron annihilation gamma rays and cosmic ray muons were used. For energy measurements, both detectors achieved an energy resolution of approximately 35% at 511 keV. For timing measurements using 511 keV, coincidence time resolutions (CTRs) of approximately 50 ps full width at half maximum (FWHM) were obtained for both detectors. In contrast, for cosmic ray muon experiments where cosmic ray muon energy is typically on the order of GeV, CTRs were measured to be 25.1 and 16.8 ps FWHM for barium fluoride- and bismuth germanate-based detectors, respectively. The versatile scintillator-integration technique established in this study can broaden the applicability of the newly developed SCI-IMPs. In particular, these results demonstrate that the developed detectors push the boundaries of timing performance while retaining energy measurement and hold promise for future applications in fundamental physics experiments and medical imaging.

physics.ins-det↗

arXiv2025

Unifying Agent Interaction and World Information for Multi-agent Coordination

This work presents a novel representation learning framework, *interaction-world* latent (IWoL), to facilitate *team coordination* in multi-agent reinforcement learning (MARL). Building effective representation for team coordination is a challenging problem, due to the intricate dynamics emerging from multi-agent interaction and incomplete information induced by local observations. Our key insight is to construct a learnable representation space that jointly captures inter-agent relations and task-specific world information by directly modeling communication protocols. This representation enables fully decentralized execution with implicit coordination while avoiding the drawbacks of explicit message passing, for example, slower decision-making, vulnerability to malicious attackers, and sensitivity to bandwidth limitations. In practice, our representation can be used not only as an implicit latent for each agent, but also as an explicit message for communication. Across four challenging MARL benchmarks, we evaluate both variants and show that IWoL provides a simple yet powerful key for team coordination. Moreover, we demonstrate that our representation can be combined with existing MARL algorithms to further enhance their performance.

cs.AI↗

arXiv2025

Policy Compatible Skill Incremental Learning via Lazy Learning Interface

Skill Incremental Learning (SIL) is the process by which an embodied agent expands and refines its skill set over time by leveraging experience gained through interaction with its environment or by the integration of additional data. SIL facilitates efficient acquisition of hierarchical policies grounded in reusable skills for downstream tasks. However, as the skill repertoire evolves, it can disrupt compatibility with existing skill-based policies, limiting their reusability and generalization. In this work, we propose SIL-C, a novel framework that ensures skill-policy compatibility, allowing improvements in incrementally learned skills to enhance the performance of downstream policies without requiring policy re-training or structural adaptation. SIL-C employs a bilateral lazy learning-based mapping technique to dynamically align the subtask space referenced by policies with the skill space decoded into agent behaviors. This enables each subtask, derived from the policy's decomposition of a complex task, to be executed by selecting an appropriate skill based on trajectory distribution similarity. We evaluate SIL-C across diverse SIL scenarios and demonstrate that it maintains compatibility between evolving skills and downstream policies while ensuring efficiency throughout the learning process.

cs.LG↗

arXiv2025

NeSyC: A Neuro-symbolic Continual Learner For Complex Embodied Tasks In Open Domains

We explore neuro-symbolic approaches to generalize actionable knowledge, enabling embodied agents to tackle complex tasks more effectively in open-domain environments. A key challenge for embodied agents is the generalization of knowledge across diverse environments and situations, as limited experiences often confine them to their prior knowledge. To address this issue, we introduce a novel framework, NeSyC, a neuro-symbolic continual learner that emulates the hypothetico-deductive model by continually formulating and validating knowledge from limited experiences through the combined use of Large Language Models (LLMs) and symbolic tools. Specifically, we devise a contrastive generality improvement scheme within NeSyC, which iteratively generates hypotheses using LLMs and conducts contrastive validation via symbolic tools. This scheme reinforces the justification for admissible actions while minimizing the inference of inadmissible ones. Additionally, we incorporate a memory-based monitoring scheme that efficiently detects action errors and triggers the knowledge refinement process across domains. Experiments conducted on diverse embodied task benchmarks-including ALFWorld, VirtualHome, Minecraft, RLBench, and a real-world robotic scenario-demonstrate that NeSyC is highly effective in solving complex embodied tasks across a range of open-domain environments.

cs.AI↗

arXiv2024

Incremental Learning of Retrievable Skills For Efficient Continual Task Adaptation

Continual Imitation Learning (CiL) involves extracting and accumulating task knowledge from demonstrations across multiple stages and tasks to achieve a multi-task policy. With recent advancements in foundation models, there has been a growing interest in adapter-based CiL approaches, where adapters are established parameter-efficiently for tasks newly demonstrated. While these approaches isolate parameters for specific tasks and tend to mitigate catastrophic forgetting, they limit knowledge sharing among different demonstrations. We introduce IsCiL, an adapter-based CiL framework that addresses this limitation of knowledge sharing by incrementally learning shareable skills from different demonstrations, thus enabling sample-efficient task adaptation using the skills particularly in non-stationary CiL environments. In IsCiL, demonstrations are mapped into the state embedding space, where proper skills can be retrieved upon input states through prototype-based memory. These retrievable skills are incrementally learned on their corresponding adapters. Our CiL experiments with complex tasks in Franka-Kitchen and Meta-World demonstrate robust performance of IsCiL in both task adaptation and sample-efficiency. We also show a simple extension of IsCiL for task unlearning scenarios.

cs.LG↗

arXiv2024

Segmented readout for Cherenkov time-of-flight positron emission tomography detectors based on bismuth germanate

Positron emission tomography (PET) is the most sensitive biomedical imaging modality for non-invasively detecting and visualizing positron-emitting radiopharmaceuticals within a subject. In PET, measuring the time-of-flight (TOF) information for each pair of 511-keV annihilation photons improves effective sensitivity but requires high timing resolution. Hybrid materials that emit both scintillation and Cherenkov photons, such as bismuth germanate (BGO), recently offer the potential for more precise timing information from Cherenkov photons while maintaining adequate energy resolution from scintillation photons. However, a significant challenge in using such hybrid materials for TOF PET applications lies in the event-dependent timing spread caused by the mixed detection of Cherenkov and scintillation photons due to relatively lower production of Cherenkov photons. This study introduces an innovative approach by segmenting silicon photomultiplier (SiPM) pixels coupled to a single crystal, rather than using traditional SiPMs that are as large as or larger than the crystals they read. We demonstrated that multiple time stamps and photon counts obtained from the segmented SiPM can classify events by providing temporal photon density, effectively addressing this challenge. The approach and findings would lead to new opportunities in applications that require precise timing and photon counting, spanning the fields of medical imaging, high-energy physics, and optical physics.

physics.ins-det↗

arXiv2024

FPGA-based digitizer for BGO-based time-of-flight PET

We present a novel FPGA-based bismuth germanate (BGO) time-of-flight (TOF) digitizer, implemented on an FPGA (VC707 evaluation kit, Xilinx). This digitizer was designed to address the recently highlighted characteristics of BGO, which generates scintillation and prompt Cerenkov photons when 511-keV photon interacts with BGO. The developed digitizer independently processes these two types of photons for precise energy and timing measurements. The digitizer incorporates a noise-resistant binary counter that measures energy signals using a time-over-threshold (TOT) method. For timing measurement, we employ an embedded dual-side monitoring time-to-digital converter (TDC), which efficiently captures timing information while maintaining low resource usage. We validated the efficacy of our FPGA-based TOF digitizer through extensive experiments, including both an electrical setup and a coincidence test using BGO pixels. Our evaluations of TOT energy and timing performance utilized two $3 \times 3 \times 20 \ \text{mm}^3$ BGO pixels coupled to CHK-HD MT silicon photomultipliers (SiPMs). The digitizer achieved a coincidence timing resolution (CTR) of 407 ps full width at half maximum (FWHM) for coincidence events falling within the full width at tenth maximum (FWTM) of the measured TOT energy spectrum. Notably, when measured with an oscilloscope, the same detector pair exhibited a CTR of 403 ps FWHM, confirming that the performance of the developed digitizer is comparable to that of oscilloscopes. With the low resource usage of our design, it offers significant potential for scalability, making it particularly promising for multi-channel BGO-based PET systems.

physics.ins-det↗

arXiv2024

Cross-Detection and Dual-Side Monitoring Schemes for FPGA-Based High-Accuracy and High-Precision Time-to-Digital Converters

This study presents a novel field-programmable gate array (FPGA)-based Time-to-Digital Converter (TDC) design suitable for high timing resolution applications, utilizing two new techniques. First, a cross-detection (CD) method is introduced that minimizes the occurrence of bubbles, which cause inaccuracy in the timing measurement of a TDC in thermometer codes, by altering the conventional sampling pattern, thereby yielding an average bin size half of its typical size. The second technique employs dual-side monitoring (DSM) of thermometer codes, including end-of-propagation (EOP) and start-of-propagation (SOP). Distinct from conventional TDCs, which focus solely on SOP thermometer codes, this technique utilizes EOP to calibrate SOP, simultaneously enhancing time resolution and the TDC's stability against changes in temperature and location. The proposed DSM scheme necessitates only an additional CARRY4 for capturing the EOP thermometer code, rendering it a resource-efficient solution. The CD-DSM TDC has been successfully implemented on a Virtex-7 Xilinx FPGA (a 28-nm process), with an average bin size of 6.1 ps and a root mean square of 3.8 ps. Compared to conventional TDCs, the CD-DSM TDC offers superior linearity. The successful measurement of ultra-high coincidence timing resolution (CTR) from two Cerenkov radiator integrated microchannel plate photomultiplier tubes (CRI-MCP-PMTs) was conducted with the CD-DSM TDCs for sub-100 ps timing measurements. A comparison with current-edge TDCs further highlights the superior performance of the CD-DSM TDCs.

physics.ins-det↗

arXiv2024

One-shot Imitation in a Non-Stationary Environment via Multi-Modal Skill

One-shot imitation is to learn a new task from a single demonstration, yet it is a challenging problem to adopt it for complex tasks with the high domain diversity inherent in a non-stationary environment. To tackle the problem, we explore the compositionality of complex tasks, and present a novel skill-based imitation learning framework enabling one-shot imitation and zero-shot adaptation; from a single demonstration for a complex unseen task, a semantic skill sequence is inferred and then each skill in the sequence is converted into an action sequence optimized for environmental hidden dynamics that can vary over time. Specifically, we leverage a vision-language model to learn a semantic skill set from offline video datasets, where each skill is represented on the vision-language embedding space, and adapt meta-learning with dynamics inference to enable zero-shot skill adaptation. We evaluate our framework with various one-shot imitation scenarios for extended multi-stage Meta-world tasks, showing its superiority in learning complex tasks, generalizing to dynamics changes, and extending to different demonstration conditions and modalities, compared to other baselines.

cs.AI↗