The center studies the complete path from multimodal representation to cognition, generation, interaction, and real-world application.中心关注从多模态表征到认知理解、内容生成、交互决策与真实场景应用的完整研究链条。
We explore unified representations across text, vision, audio, motion, and 3D modalities, with emphasis on cross-modal alignment, complex reasoning, context learning, and interpretable model behavior.探索文本、视觉、音频、运动与三维等模态的统一表征，重点研究跨模态对齐、复杂推理、上下文学习及模型行为解释。
We investigate perception and interaction in physical environments, including human and robotic representation learning, sensorimotor control, motion understanding, physical commonsense, and recovery of 3D behavior under real-world conditions.研究智能体在物理环境中的感知与交互，包括人类及机器人表征学习、感知运动控制、动作理解、物理常识，以及真实复杂环境中的三维行为恢复。
We study multimodal representation and reasoning for healthcare, including medical image retrieval and segmentation, cross-modal evidence integration, and cognitive decision support for clinically meaningful applications.研究面向医疗健康的多模态表征与推理，包括医学图像检索与分割、跨模态证据融合，以及服务临床应用的认知决策支持。
We develop generative intelligence for high-fidelity and controllable content, covering image editing, dynamic video, 3D scene generation and editing, human motion, and interactive virtual humans.面向高保真、可控制的生成智能，研究图像编辑、动态视频、三维场景生成与编辑、人类动作及交互式虚拟人。
Slides converted from the embodied intelligence portfolio, showing the center's work across digital humans, virtual humans, embodied agents, and cross-space cognitive applications.由具身智能 PPT 转换生成，展示中心围绕数字人、虚拟人、具身人以及跨空间认知应用形成的研究布局。
A concise overview distilled from the center's technical portfolio in efficient vision, infrared remote sensing, and multimodal understanding.从中心在高效视觉与多模态理解方向的技术积累中，提炼形成清晰可展示的研究与应用链条。
CAS-ViT introduces convolutional additive self-attention for lightweight, high-throughput perception. The backbone has been applied in multiple SenseTime embedded terminals.以 CAS-ViT 为代表的高效视觉骨干，通过卷积加性自注意力降低端侧推理成本，已应用于商汤科技多个嵌入式终端。
The center focuses on foundational research in pose-language pretraining, unified embeddings, long-video compression, and multimodal evaluation, serving advanced international industry needs.围绕姿态—语言预训练、统一嵌入、长视频关键帧压缩与多模态评测开展基础研究，面向海外企业对复杂视频图文理解的应用需求。
Physical-prior completion, CAS-ViT efficient representation, 3D scene vectorization, and language-guided editing form an end-to-end technical path.物理先验补全、CAS-ViT 高效表征、三维场景向量化与语言引导编辑，构成从感知到端侧部署的技术链条。
Large-scale pose-language pretraining, unified multimodal embeddings, joint prediction, and long-sequence compression support transferable multimodal understanding.大规模姿态—语言预训练、多模态统一嵌入、联合预测与长序列压缩，支撑可迁移的视频图像文本理解能力。
