arXiv subjects

Kyumin Choi

Publications and source records attributed to Kyumin Choi.

7 recordsLinked to original sources

arXiv2026

FastOPD: On-Policy Distillation for Lightweight VLA Deployment

Vision-Language-Action (VLA) foundation models have scaled rapidly to enhance manipulation performance and generalizability, but this scaling incurs high computational costs that render real-world deployment increasingly challenging. Existing approaches typically mitigate this issue by designing smaller architectures or reducing the iterative denoising steps in flow-based policies. In this work, we propose FastOPD, a foundation-to-lightweight VLA framework that enables the practical deployment of large-scale VLAs through efficient on-policy distillation. Specifically, FastOPD adapts a flow map for single-state teacher supervision and combines it with a self-consistency objective to construct a compact student that learns the teacher dynamics. Furthermore, we theoretically demonstrate that minimizing this objective allows the distilled student to recover a distribution on par with that induced by an ideal few-step teacher model. We evaluate FastOPD across diverse foundation policies in simulation and real-world experiments. On LIBERO, FastOPD retains 84% of the performance of $π_{0.5}$ with only two inference steps, reducing inference latency by 78.1% while outperforming existing few-step distillation baselines in average success rate. With LingBot-VLA as the teacher, FastOPD improves the single-step success rate over the base student by 15.9 percentage points on RoboTwin 2.0. We further demonstrate its applicability to a World Action Model (WAM) and deploy a compact student distilled from MolmoAct2 on a real robot.

cs.RO↗

arXiv2026

Don't Throw Away the Tail: Action Upcycling for Policy Acceleration

Modern robot policies predict a chunk of future actions from a single observation, execute only a prefix, and discard the rest before replanning. Choosing the length of this prefix, the execution horizon, poses a trade-off between reactivity and efficiency. A short horizon keeps the policy reactive to the environment, but requires frequent policy calls. Recent test-time methods adaptively select the horizon for each chunk, but they either read model internals, where the signal must be chosen for each architecture, or draw extra samples, which adds cost. We propose Action Upcycling, a training-free algorithm that reuses actions the policy would otherwise discard, without accessing model internals or drawing extra samples. We find that discarded actions stay close to their replanned versions as long as the action velocity remains smooth. Action Upcycling therefore extends the execution horizon up to the point where the velocity begins to fluctuate. Extensive experiments on simulated and real-world manipulation tasks show that Action Upcycling reduces policy calls by 1.2-1.7x with no loss in success rate, across multiple Vision-Language-Action Models (VLAs) and even a World Action Model (WAM). It applies to any chunked policy at negligible cost and is orthogonal to other policy acceleration methods such as few-step sampling and streaming action decoding, opening a new axis for policy acceleration.

cs.RO↗

arXiv2026

DriftOPD: Sequence-Level Reverse-KL Distillation for One-Step VLA Policies

Vision-Language-Action (VLA) models increasingly rely on action experts that generate short action chunks under receding-horizon control. While chunk-level training is convenient across robot embodiments, it optimizes local action likelihood without explicitly accounting for long-horizon task success. Sequence-level reinforcement learning can address this limitation, but typically requires policy rollouts and closed-loop interaction, which are costly for real-robot manipulation. We introduce DriftOPD, a teacher-free, rollout-free framework for sequence-level on-policy distillation of continuous VLA action experts. We show that the sequence-level reverse Kullback-Leibler (KL) divergence decomposes into a chunk-level reverse-KL term and a future-potential term that captures the long-horizon effect of the current action. DriftOPD optimizes these two terms using a one-step drifting objective and a Q-function critic learned from offline demonstrations, respectively, enabling sequence-level optimization with only offline data and one-step action generation. Across multiple VLA architectures in simulation and real-world manipulation, DriftOPD generally outperforms existing one-step distillation baselines while achieving task success performance comparable to multi-step teacher policies. These results demonstrate that long-horizon behavior can be effectively distilled into one-step VLA action experts without online interaction or a separate teacher.

cs.RO↗

arXiv2026

Adjoint Guidance Flow: Amortized Critic Guidance for VLA Policies

Flow-based Vision-Language-Action (VLA) policies are typically trained by behavior cloning and thus do not explicitly optimize long-term task return. Critic guidance steers generation toward higher-value actions, but existing methods differentiate the critic through a one-step surrogate of the sampler and back-propagate a critic ensemble at every flow step. In contrast, here we propose Adjoint Guidance Flow (AGF), which amortizes trajectory-aware critic guidance into a lightweight guidance network while preserving the pretrained VLA policy. Specifically, we formulate critic-guided flow generation as a deterministic optimal control problem, whose optimal guidance is a costate that carries the terminal critic gradient back through the remaining flow, and regress the guidance network onto this costate while keeping both the VLA and critic frozen. This design provides favorable memory and throughput scaling during training, and inference needs one guidance-network forward pass per step, without the critic ensemble, back-propagation, or adjoint computation. Across LIBERO, RoboCasa, and LIBERO-Pro, AGF consistently improves pretrained VLAs, remains competitive with critic-guidance and policy-fine-tuning baselines, and is the most robust method when a single guidance strength is deployed across tasks. Compared with QGF, AGF runs $3.6\times$ faster per guidance step with $7.0\times$ fewer parameters, with comparable and even better performance, showing that critic guidance can be trajectory-aware and lightweight.

cs.LG↗

arXiv2026

RAPID: Layer-Wise Redundancy-Aware Pruning and Importance-Driven Token Merging for Efficient ViT

Vision Transformers (ViTs) achieve strong performance but suffer from high computational costs due to quadratic self-attention complexity. Although token reduction techniques such as pruning and merging mitigate this, they typically overlook how representations evolve across network depth. We propose RAPID, a depth-aware token reduction framework that adapts reduction strategies to the layer-wise characteristics of token representations. The primary methodological contribution is a bifurcated strategy: in shallow-to-middle layers, RAPID employs a redundancy-similarity aware pruning metric to eliminate over-represented local patterns. As features transition to global semantic concepts in deeper layers, the framework shifts to an importance-similarity aware merging mechanism. This stage leverages classification (CLS) token attention weights to protect semantically critical tokens while fusing less important but similar neighbors. Empirical validation on ImageNet-1K using ViT and DeiT architectures demonstrates that RAPID establishes a superior accuracy-compression Pareto frontier compared to plug-and-play baselines such as ToMe and ToFu. RAPID is particularly robust in aggressive compression regimes, achieving up to 4.29% higher accuracy than ToMe at extreme reduction rates. Our framework provides a training-free template for optimizing vision models by aligning reduction strategies with hierarchical feature evolution.

cs.CV↗

arXiv2026

Conditional Unbalanced Optimal Transport Maps: An Outlier-Robust Framework for Conditional Generative Modeling

Conditional Optimal Transport (COT) problem aims to find a transport map between conditional source and target distributions while minimizing the transport cost. Recently, these transport maps have been utilized in conditional generative modeling tasks to establish efficient mappings between the distributions. However, classical COT inherits a fundamental limitation of optimal transport, i.e., sensitivity to outliers, which arises from the hard distribution matching constraints. This limitation becomes more pronounced in a conditional setting, where each conditional distribution is estimated from a limited subset of data. To address this, we introduce the Conditional Unbalanced Optimal Transport (CUOT) framework, which relaxes conditional distribution-matching constraints through Csiszár divergence penalties while strictly preserving the conditioning marginals. We establish a rigorous formulation of the CUOT problem and derive its dual and semi-dual formulations. Based on the semi-dual form, we propose Conditional Unbalanced Optimal Transport Maps (CUOTM), an outlier-robust conditional generative model built upon a triangular $c$-transform parameterization. We theoretically justify the validity of this parameterization by proving that the optimal triangular map satisfies the $c$-transform relationships. Our experiments on 2D synthetic and image-scale datasets demonstrate that CUOTM achieves superior outlier robustness and competitive distribution-matching performance compared to existing COT-based baselines, while maintaining high sampling efficiency.

cs.LG↗

arXiv2019

Using offline metrics and user behavior analysis to combine multiple systems for music recommendation

There are many offline metrics that can be used as a reference for evaluation and optimization of the performance of recommender systems. Hybrid recommendation approaches are commonly used to improve some of those metrics by combining different systems. In this work we focus on music recommendation and propose a new way to improve recommendations, with respect to a desired metric of choice, by combining multiple systems for each user individually based on their expected performance. Essentially, our approach consists in predicting an expected error that each system will produce for each user based on their previous activity. To this end, we propose to train regression models for different metrics predicting the performance of each system based on a number of features characterizing previous user behavior in the system. We then use different fusion strategies to combine recommendations generated by each system. Following this approach one can optimize the final hybrid system with respect to the desired metric of choice. As a proof of concept, we conduct experiments combining two recommendation systems, a Matrix Factorization model and a popularity-based recommender. We use the data provided by Melon, a Korean music streaming service, to train and evaluate the performance of the systems.

cs.IR↗