Selected work

Representative Papers

UniPhysGen: Unified Physical Grounding for Simulation-Ready 3D Assets

Jul 15, 2026

Existing 3D assets commonly lack consistent physical semantics—such as articulated motion and intrinsic physical properties—hindering their use in high-fidelity robotic simulation. This work proposes UniPhys, a novel framework that, for the first time, enables unified physical semantic modeling of heterogeneous 3D assets. UniPhys mitigates geometric shortcut biases arising from part decomposition discrepancies through a geometry-robust joint grounding mechanism and employs a joint modeling architecture to end-to-end infer both articulation semantics and physical attributes. Leveraging this approach, we introduce UniPhys-40K, a large-scale dataset, and UniPhys-Bench, a comprehensive benchmark. We also develop UniPhysGen, which achieves state-of-the-art performance in joint grounding and physical property estimation, producing assets directly deployable in robotic simulators for realistic physical interactions.

0 citationsRead paper

SPEAR: A Simulator for Photorealistic Embodied AI Research

Jul 07, 2026

Existing high-fidelity simulators struggle to meet the demands of embodied intelligence research in terms of generality, programmability, and rendering efficiency. To address this gap, this work proposes SPEAR—a Python-based simulation library built on Unreal Engine—that exposes over 14,000 UE functions to Python for the first time, enabling unprecedented procedural control. SPEAR employs a modular plugin architecture and a high-level task graph scheduling model to support deterministic execution of complex tasks, delivering high-definition multimodal images along with accurate annotations to NumPy arrays at an efficient rate of 73 FPS. The system has been successfully applied to diverse tasks including agent control, city-scale rendering, multi-view facial imaging, physics-informed co-simulation, and language-driven scene editing, significantly enhancing both simulation efficiency and functional coverage.

0 citationsRead paper

Controllable Texture Tiling with Transformed RoPE-Enhanced Diffusion Models

Jun 22, 2026

This work proposes a controllable, high-fidelity texture tiling framework based on diffusion transformers to address the limitations of existing methods, which often fail to precisely regulate the frequency, orientation, and scale of repeating patterns while preserving reference texture structure and scene-consistent lighting and geometry. The approach decouples spatial manipulation from content generation, enabling accurate tiling according to user-specified parameters. Key innovations include a coordinate-transformed Rotary Position Embedding (RoPE) that applies affine transformations to relative position embeddings for direct control over tiling attributes, and a disentangled attention mask designed to prevent semantic leakage and maintain structural integrity. Experiments demonstrate that the method surpasses current state-of-the-art techniques in both control precision and texture fidelity, while effectively harmonizing reference textures with scene-specific illumination and geometric cues.

0 citationsRead paper

Rein3D: Reinforced 3D Indoor Scene Generation with Panoramic Video Diffusion Models

Apr 12, 2026

This work addresses the challenges of large-scale missing regions and global inconsistency in 3D indoor scene reconstruction from sparse inputs. The authors propose a “reconstruction–optimization” paradigm: an initial panoramic video is first generated via radial exploration, then refined through a temporally consistent panoramic video-to-video diffusion model for inpainting and super-resolution. The enhanced video serves as a pseudo-ground-truth to guide the global optimization of a 3D Gaussian Splatting field. This study introduces the first coupling of panoramic video diffusion priors with 3D Gaussian representations, presenting a novel video-guided 3D optimization mechanism and releasing PanoV2V-15K, the first large-scale dataset for panoramic video inpainting. The method significantly outperforms existing baselines in long-range exploration tasks, producing photorealistic and globally consistent 360-degree indoor scenes.

0 citationsRead paper

Arcadia: Toward a Full-Lifecycle Framework for Embodied Lifelong Learning

Nov 25, 2025

Existing embodied lifelong learning systems typically optimize isolated components—such as data collection or deployment—independently, hindering sustained improvement and cross-environment generalization. This paper introduces Arcadia, the first holistic framework that models embodied learning as an indivisible closed-loop lifecycle, encompassing four tightly coupled stages: autonomous exploration, generative scene reconstruction, shared multimodal representation learning, and simulation-driven evolution. Its key innovations include establishing the first sim-from-real feedback loop between physical and virtual domains, unifying self-evolving exploration, generative data augmentation, and a unified multimodal representation architecture. Arcadia enables reproducible, cross-task and cross-environment evaluation. Empirically, it achieves continuous performance gains on navigation and manipulation benchmarks and successfully transfers learned policies to real-world robots, demonstrating robustness and generalizability.

0 citationsRead paper

Recent publications

Latest Papers

UniPhysGen: Unified Physical Grounding for Simulation-Ready 3D Assets

Jul 15, 2026

Existing 3D assets commonly lack consistent physical semantics—such as articulated motion and intrinsic physical properties—hindering their use in high-fidelity robotic simulation. This work proposes UniPhys, a novel framework that, for the first time, enables unified physical semantic modeling of heterogeneous 3D assets. UniPhys mitigates geometric shortcut biases arising from part decomposition discrepancies through a geometry-robust joint grounding mechanism and employs a joint modeling architecture to end-to-end infer both articulation semantics and physical attributes. Leveraging this approach, we introduce UniPhys-40K, a large-scale dataset, and UniPhys-Bench, a comprehensive benchmark. We also develop UniPhysGen, which achieves state-of-the-art performance in joint grounding and physical property estimation, producing assets directly deployable in robotic simulators for realistic physical interactions.

0 citationsRead paper

SPEAR: A Simulator for Photorealistic Embodied AI Research

Jul 07, 2026

Existing high-fidelity simulators struggle to meet the demands of embodied intelligence research in terms of generality, programmability, and rendering efficiency. To address this gap, this work proposes SPEAR—a Python-based simulation library built on Unreal Engine—that exposes over 14,000 UE functions to Python for the first time, enabling unprecedented procedural control. SPEAR employs a modular plugin architecture and a high-level task graph scheduling model to support deterministic execution of complex tasks, delivering high-definition multimodal images along with accurate annotations to NumPy arrays at an efficient rate of 73 FPS. The system has been successfully applied to diverse tasks including agent control, city-scale rendering, multi-view facial imaging, physics-informed co-simulation, and language-driven scene editing, significantly enhancing both simulation efficiency and functional coverage.

0 citationsRead paper

Controllable Texture Tiling with Transformed RoPE-Enhanced Diffusion Models

Jun 22, 2026

This work proposes a controllable, high-fidelity texture tiling framework based on diffusion transformers to address the limitations of existing methods, which often fail to precisely regulate the frequency, orientation, and scale of repeating patterns while preserving reference texture structure and scene-consistent lighting and geometry. The approach decouples spatial manipulation from content generation, enabling accurate tiling according to user-specified parameters. Key innovations include a coordinate-transformed Rotary Position Embedding (RoPE) that applies affine transformations to relative position embeddings for direct control over tiling attributes, and a disentangled attention mask designed to prevent semantic leakage and maintain structural integrity. Experiments demonstrate that the method surpasses current state-of-the-art techniques in both control precision and texture fidelity, while effectively harmonizing reference textures with scene-specific illumination and geometric cues.

0 citationsRead paper

Rein3D: Reinforced 3D Indoor Scene Generation with Panoramic Video Diffusion Models

Apr 12, 2026

This work addresses the challenges of large-scale missing regions and global inconsistency in 3D indoor scene reconstruction from sparse inputs. The authors propose a “reconstruction–optimization” paradigm: an initial panoramic video is first generated via radial exploration, then refined through a temporally consistent panoramic video-to-video diffusion model for inpainting and super-resolution. The enhanced video serves as a pseudo-ground-truth to guide the global optimization of a 3D Gaussian Splatting field. This study introduces the first coupling of panoramic video diffusion priors with 3D Gaussian representations, presenting a novel video-guided 3D optimization mechanism and releasing PanoV2V-15K, the first large-scale dataset for panoramic video inpainting. The method significantly outperforms existing baselines in long-range exploration tasks, producing photorealistic and globally consistent 360-degree indoor scenes.

0 citationsRead paper

Arcadia: Toward a Full-Lifecycle Framework for Embodied Lifelong Learning

Nov 25, 2025

Existing embodied lifelong learning systems typically optimize isolated components—such as data collection or deployment—independently, hindering sustained improvement and cross-environment generalization. This paper introduces Arcadia, the first holistic framework that models embodied learning as an indivisible closed-loop lifecycle, encompassing four tightly coupled stages: autonomous exploration, generative scene reconstruction, shared multimodal representation learning, and simulation-driven evolution. Its key innovations include establishing the first sim-from-real feedback loop between physical and virtual domains, unifying self-evolving exploration, generative data augmentation, and a unified multimodal representation architecture. Arcadia enables reproducible, cross-task and cross-environment evaluation. Empirically, it achieves continuous performance gains on navigation and manipulation benchmarks and successfully transfers learned policies to real-world robots, demonstrating robustness and generalizability.

0 citationsRead paper