Selected work

Representative Papers

Enhancing Human-Likeness in Reinforcement Learning Agents via Hierarchical Macro Action Quantization

May 29, 2026

Reinforcement learning agents often lack interpretability and reliability due to behavioral discrepancies from humans. To address this, this work proposes Hierarchical Macro-Action Quantization (HiMAQ), a method that encodes human demonstrations into structured behavioral units through a two-level vector quantization mechanism: first clustering fine-grained sub-actions and then aggregating them into high-level macro-actions. Experiments on the D4RL benchmark demonstrate that HiMAQ consistently enhances the human-likeness of agent behavior across multiple offline reinforcement learning algorithms—including IQL, SAC, and RLPD—while maintaining comparable or higher task success rates. HiMAQ outperforms its non-hierarchical counterpart, MAQ, exhibiting strong generalization capability and practical utility.

0 citationsRead paper

Unsupervised Skeleton-Based Action Segmentation via Hierarchical Spatiotemporal Vector Quantization

Apr 16, 2026

This work addresses the challenges of segment length bias and insufficient spatiotemporal modeling in unsupervised skeleton-based action segmentation by proposing a hierarchical spatiotemporal vector quantization framework. The method employs a two-stage vector quantization process: first mapping raw skeleton sequences into fine-grained sub-action units, then aggregating these units into action-level representations. By jointly reconstructing both the skeleton data and their corresponding timestamps, the model enables end-to-end unsupervised spatiotemporal learning. To the best of our knowledge, this is the first effort to introduce hierarchical vector quantization to this task, effectively integrating spatial and temporal cues and substantially mitigating segment length bias. The approach achieves state-of-the-art performance across multiple benchmarks—including HuGaDB, LARa, and BABEL—significantly outperforming non-hierarchical baseline methods.

0 citationsRead paper

A Hierarchical Spatiotemporal Action Tokenizer for In-Context Imitation Learning in Robotics

Apr 16, 2026

This work addresses the limitation in contextual imitation learning where action representations lack effective spatiotemporal structure modeling. To this end, the authors propose the Hierarchical Spatio-Temporal Action Tokenizer (HiST-AT), which introduces, for the first time, a hierarchical spatio-temporal modeling mechanism that jointly leverages spatial and temporal cues during action tokenization. HiST-AT employs a two-level vector quantization scheme to simultaneously capture fine-grained and coarse-grained action clusters while reconstructing both actions and their corresponding timestamps. Experimental results demonstrate that HiST-AT achieves state-of-the-art performance across multiple simulated and real-world robotic manipulation benchmarks, significantly enhancing the efficiency, generalization, and robustness of imitation learning.

0 citationsRead paper

Procedure Learning via Regularized Gromov-Wasserstein Optimal Transport

Jul 21, 2025

This work addresses the self-supervised discovery and temporal modeling of key steps in unlabeled procedural videos, tackling challenges including action repetition, background clutter, and step-order variability. We propose a Gromov-Wasserstein (GW) optimal transport framework augmented with structural priors: a contrastive regularization term is introduced to constrain the embedding space, preventing degenerate solutions and enhancing robustness in cross-video temporal alignment and key-step identification. Our method unifies self-supervised learning, inter-frame mapping modeling, and structural-aware GW distance computation. Evaluated on EgoProceL, ProceL, and CrossTask benchmarks, it significantly outperforms prior approaches—including OPEL—achieving state-of-the-art performance. Notably, this is the first work to explicitly incorporate structural priors into the GW optimization process. The resulting framework delivers an interpretable, robust paradigm for temporal modeling in procedural video understanding.

0 citationsRead paper

Joint Self-Supervised Video Alignment and Action Segmentation

Mar 21, 2025

This paper addresses the joint modeling challenge of unsupervised video alignment and action segmentation. We propose the first unified self-supervised learning framework for both tasks. Methodologically, we formulate them jointly as a single optimal transport problem, introducing a structural-prior-augmented Gromov–Wasserstein (GW) metric to simultaneously optimize alignment and segmentation within a single model—without any manual annotations. Our core contributions are: (1) a multi-task unified objective function based on a structure-aware GW distance; and (2) a differentiable self-supervised optimization mechanism enabling end-to-end joint training. On multiple benchmarks, our method achieves state-of-the-art performance in video alignment and significantly outperforms existing unsupervised approaches in action segmentation mAP. Moreover, it reduces GPU memory consumption and storage overhead by approximately 50%.

0 citationsRead paper

Recent publications

Latest Papers

Enhancing Human-Likeness in Reinforcement Learning Agents via Hierarchical Macro Action Quantization

May 29, 2026

Reinforcement learning agents often lack interpretability and reliability due to behavioral discrepancies from humans. To address this, this work proposes Hierarchical Macro-Action Quantization (HiMAQ), a method that encodes human demonstrations into structured behavioral units through a two-level vector quantization mechanism: first clustering fine-grained sub-actions and then aggregating them into high-level macro-actions. Experiments on the D4RL benchmark demonstrate that HiMAQ consistently enhances the human-likeness of agent behavior across multiple offline reinforcement learning algorithms—including IQL, SAC, and RLPD—while maintaining comparable or higher task success rates. HiMAQ outperforms its non-hierarchical counterpart, MAQ, exhibiting strong generalization capability and practical utility.

0 citationsRead paper

Unsupervised Skeleton-Based Action Segmentation via Hierarchical Spatiotemporal Vector Quantization

Apr 16, 2026

This work addresses the challenges of segment length bias and insufficient spatiotemporal modeling in unsupervised skeleton-based action segmentation by proposing a hierarchical spatiotemporal vector quantization framework. The method employs a two-stage vector quantization process: first mapping raw skeleton sequences into fine-grained sub-action units, then aggregating these units into action-level representations. By jointly reconstructing both the skeleton data and their corresponding timestamps, the model enables end-to-end unsupervised spatiotemporal learning. To the best of our knowledge, this is the first effort to introduce hierarchical vector quantization to this task, effectively integrating spatial and temporal cues and substantially mitigating segment length bias. The approach achieves state-of-the-art performance across multiple benchmarks—including HuGaDB, LARa, and BABEL—significantly outperforming non-hierarchical baseline methods.

0 citationsRead paper

A Hierarchical Spatiotemporal Action Tokenizer for In-Context Imitation Learning in Robotics

Apr 16, 2026

This work addresses the limitation in contextual imitation learning where action representations lack effective spatiotemporal structure modeling. To this end, the authors propose the Hierarchical Spatio-Temporal Action Tokenizer (HiST-AT), which introduces, for the first time, a hierarchical spatio-temporal modeling mechanism that jointly leverages spatial and temporal cues during action tokenization. HiST-AT employs a two-level vector quantization scheme to simultaneously capture fine-grained and coarse-grained action clusters while reconstructing both actions and their corresponding timestamps. Experimental results demonstrate that HiST-AT achieves state-of-the-art performance across multiple simulated and real-world robotic manipulation benchmarks, significantly enhancing the efficiency, generalization, and robustness of imitation learning.

0 citationsRead paper

Procedure Learning via Regularized Gromov-Wasserstein Optimal Transport

Jul 21, 2025

This work addresses the self-supervised discovery and temporal modeling of key steps in unlabeled procedural videos, tackling challenges including action repetition, background clutter, and step-order variability. We propose a Gromov-Wasserstein (GW) optimal transport framework augmented with structural priors: a contrastive regularization term is introduced to constrain the embedding space, preventing degenerate solutions and enhancing robustness in cross-video temporal alignment and key-step identification. Our method unifies self-supervised learning, inter-frame mapping modeling, and structural-aware GW distance computation. Evaluated on EgoProceL, ProceL, and CrossTask benchmarks, it significantly outperforms prior approaches—including OPEL—achieving state-of-the-art performance. Notably, this is the first work to explicitly incorporate structural priors into the GW optimization process. The resulting framework delivers an interpretable, robust paradigm for temporal modeling in procedural video understanding.

0 citationsRead paper

Joint Self-Supervised Video Alignment and Action Segmentation

Mar 21, 2025

This paper addresses the joint modeling challenge of unsupervised video alignment and action segmentation. We propose the first unified self-supervised learning framework for both tasks. Methodologically, we formulate them jointly as a single optimal transport problem, introducing a structural-prior-augmented Gromov–Wasserstein (GW) metric to simultaneously optimize alignment and segmentation within a single model—without any manual annotations. Our core contributions are: (1) a multi-task unified objective function based on a structure-aware GW distance; and (2) a differentiable self-supervised optimization mechanism enabling end-to-end joint training. On multiple benchmarks, our method achieves state-of-the-art performance in video alignment and significantly outperforms existing unsupervised approaches in action segmentation mAP. Moreover, it reduces GPU memory consumption and storage overhead by approximately 50%.

0 citationsRead paper