Selected work

Representative Papers

Quantile-Coupled Flow Matching for Distributional Reinforcement Learning

May 08, 2026

This work addresses a critical inconsistency in existing conditional flow matching (CFM) approaches for distributional reinforcement learning, where arbitrary source–target pairings yield losses misaligned with the Wasserstein distance, thereby violating the contraction property of the Bellman operator. To resolve this, the authors propose FlowIQN, which constructs quantile-aligned, monotonic optimal transport couplings by sorting source samples and Bellman targets within each minibatch, ensuring flow trajectories consistent with the Wasserstein metric. FlowIQN provides the first explicit Wasserstein projection guarantee for flow-matching distributional critics and incorporates a shortcut inference model to enhance computational efficiency. Empirical results demonstrate that FlowIQN significantly improves the Wasserstein accuracy of return distributions and achieves strong performance across multiple offline reinforcement learning benchmarks under various policy extraction settings, offering both theoretical rigor and practical effectiveness.

0 citationsRead paper

KV-Tracker: Real-Time Pose Tracking with Transformers

Dec 27, 2025

Existing multi-view geometry–based approaches for real-time 6-DoF pose tracking and online 3D reconstruction from monocular RGB video incur high computational overhead, hindering real-time performance. This paper proposes a lightweight, depth-free, object-agnostic, and scene-general framework. We introduce a novel KV caching mechanism that models global self-attention key-value pairs as compact, persistent scene representations—enabling model-agnostic, retraining-free long-term consistency modeling. Our method integrates π³ multi-view geometric priors with dynamic keyframe selection and bidirectional attention. Evaluated on benchmarks including TUM RGB-D, the framework achieves 27 FPS inference speed, up to 15× acceleration over baselines, and significantly mitigates pose drift and catastrophic forgetting.

0 citationsRead paper

4D Primitive-Mâché: Glueing Primitives for Persistent 4D Scene Reconstruction

Dec 18, 2025

This paper addresses complete and persistent 4D scene reconstruction from monocular RGB video. We propose a novel method that decomposes dynamic scenes into rigid 3D primitives and jointly optimizes their temporally consistent rigid-body motions. Our key contributions are: (1) a primitive stitching mechanism that enforces cross-frame geometric and motion consistency; and (2) an occlusion-aware motion extrapolation strategy enabling permanent object modeling and fully replayable temporal reconstruction. The framework integrates dense 2D correspondence estimation, motion clustering, primitive decomposition, and temporal consistency constraints. Evaluated on multi-object scanning and dynamic scene datasets, our approach achieves state-of-the-art performance in reconstruction completeness, geometric accuracy, and temporal replayability—outperforming existing methods across all three metrics.

0 citationsRead paper

ACE-SLAM: Scene Coordinate Regression for Neural Implicit Real-Time SLAM

Dec 15, 2025

This paper addresses the fundamental trade-off among real-time performance, memory efficiency, and relocalization accuracy in neural implicit SLAM systems. We propose the first real-time neural implicit RGB-D SLAM framework based on Scene Coordinate Regression (SCR). Our key contributions are: (i) the first integration of SCR into the neural SLAM pipeline as a lightweight, differentiable, and globally consistent implicit map representation; (ii) a compact, real-time-optimized SCR architecture enabling RGB-D multimodal feature fusion and robust mapping in dynamic environments; and (iii) end-to-end differentiable training with millisecond-scale pose optimization. Evaluated on both synthetic and real-world datasets, our method achieves state-of-the-art performance—exceeding 30 FPS, achieving relocalization latency under 5 ms, and reducing GPU memory consumption by over 60%—while ensuring low memory footprint, strong privacy preservation, and strict real-time constraints.

0 citationsRead paper

GeCCo - a Generalist Contact-Conditioned Policy for Loco-Manipulation Skills on Legged Robots

Sep 22, 2025

Existing quadrupedal locomotion methods predominantly rely on end-to-end deep reinforcement learning (DRL), requiring task-specific reward engineering for each new behavior—limiting scalability and reusability. This work introduces GeCCo (Generalized Contact-Conditioned policy), a universal low-level controller trained via DRL to robustly track arbitrary foot contact locations conditioned on real-time contact states. Unlike prior approaches, GeCCo enables plug-and-play deployment of a single low-level policy across diverse high-level tasks—including walking, obstacle negotiation, and button pressing—without retraining. By decoupling high-level task planning from low-level contact control, the framework significantly improves modularity and deployment efficiency. Experiments demonstrate strong generalization and robustness across complex terrains and multi-task scenarios, validating GeCCo’s effectiveness in bridging perception, planning, and control. This work establishes a scalable, modular paradigm for embodied locomotion control, advancing the design of adaptive and reusable legged agents.

0 citationsRead paper

Recent publications

Latest Papers

Quantile-Coupled Flow Matching for Distributional Reinforcement Learning

May 08, 2026

This work addresses a critical inconsistency in existing conditional flow matching (CFM) approaches for distributional reinforcement learning, where arbitrary source–target pairings yield losses misaligned with the Wasserstein distance, thereby violating the contraction property of the Bellman operator. To resolve this, the authors propose FlowIQN, which constructs quantile-aligned, monotonic optimal transport couplings by sorting source samples and Bellman targets within each minibatch, ensuring flow trajectories consistent with the Wasserstein metric. FlowIQN provides the first explicit Wasserstein projection guarantee for flow-matching distributional critics and incorporates a shortcut inference model to enhance computational efficiency. Empirical results demonstrate that FlowIQN significantly improves the Wasserstein accuracy of return distributions and achieves strong performance across multiple offline reinforcement learning benchmarks under various policy extraction settings, offering both theoretical rigor and practical effectiveness.

0 citationsRead paper

KV-Tracker: Real-Time Pose Tracking with Transformers

Dec 27, 2025

Existing multi-view geometry–based approaches for real-time 6-DoF pose tracking and online 3D reconstruction from monocular RGB video incur high computational overhead, hindering real-time performance. This paper proposes a lightweight, depth-free, object-agnostic, and scene-general framework. We introduce a novel KV caching mechanism that models global self-attention key-value pairs as compact, persistent scene representations—enabling model-agnostic, retraining-free long-term consistency modeling. Our method integrates π³ multi-view geometric priors with dynamic keyframe selection and bidirectional attention. Evaluated on benchmarks including TUM RGB-D, the framework achieves 27 FPS inference speed, up to 15× acceleration over baselines, and significantly mitigates pose drift and catastrophic forgetting.

0 citationsRead paper

4D Primitive-Mâché: Glueing Primitives for Persistent 4D Scene Reconstruction

Dec 18, 2025

This paper addresses complete and persistent 4D scene reconstruction from monocular RGB video. We propose a novel method that decomposes dynamic scenes into rigid 3D primitives and jointly optimizes their temporally consistent rigid-body motions. Our key contributions are: (1) a primitive stitching mechanism that enforces cross-frame geometric and motion consistency; and (2) an occlusion-aware motion extrapolation strategy enabling permanent object modeling and fully replayable temporal reconstruction. The framework integrates dense 2D correspondence estimation, motion clustering, primitive decomposition, and temporal consistency constraints. Evaluated on multi-object scanning and dynamic scene datasets, our approach achieves state-of-the-art performance in reconstruction completeness, geometric accuracy, and temporal replayability—outperforming existing methods across all three metrics.

0 citationsRead paper

ACE-SLAM: Scene Coordinate Regression for Neural Implicit Real-Time SLAM

Dec 15, 2025

This paper addresses the fundamental trade-off among real-time performance, memory efficiency, and relocalization accuracy in neural implicit SLAM systems. We propose the first real-time neural implicit RGB-D SLAM framework based on Scene Coordinate Regression (SCR). Our key contributions are: (i) the first integration of SCR into the neural SLAM pipeline as a lightweight, differentiable, and globally consistent implicit map representation; (ii) a compact, real-time-optimized SCR architecture enabling RGB-D multimodal feature fusion and robust mapping in dynamic environments; and (iii) end-to-end differentiable training with millisecond-scale pose optimization. Evaluated on both synthetic and real-world datasets, our method achieves state-of-the-art performance—exceeding 30 FPS, achieving relocalization latency under 5 ms, and reducing GPU memory consumption by over 60%—while ensuring low memory footprint, strong privacy preservation, and strict real-time constraints.

0 citationsRead paper

GeCCo - a Generalist Contact-Conditioned Policy for Loco-Manipulation Skills on Legged Robots

Sep 22, 2025

Existing quadrupedal locomotion methods predominantly rely on end-to-end deep reinforcement learning (DRL), requiring task-specific reward engineering for each new behavior—limiting scalability and reusability. This work introduces GeCCo (Generalized Contact-Conditioned policy), a universal low-level controller trained via DRL to robustly track arbitrary foot contact locations conditioned on real-time contact states. Unlike prior approaches, GeCCo enables plug-and-play deployment of a single low-level policy across diverse high-level tasks—including walking, obstacle negotiation, and button pressing—without retraining. By decoupling high-level task planning from low-level contact control, the framework significantly improves modularity and deployment efficiency. Experiments demonstrate strong generalization and robustness across complex terrains and multi-task scenarios, validating GeCCo’s effectiveness in bridging perception, planning, and control. This work establishes a scalable, modular paradigm for embodied locomotion control, advancing the design of adaptive and reusable legged agents.

0 citationsRead paper