I. Multimodal Understanding 2.1
Encoding linguistic semantics and visual signals within a shared latent space, enabling neural networks to comprehend the physical world across both spatial and temporal dimensions.
1.1
吴祥楠 Multimodal Understanding & Embodied Intelligence
Building generalist intelligent agents that perceive, reason, and act across both physical and virtual worlds.
I am a Ph.D. candidate at
UCAS
&
NLPR@CASIA,
advised by
Yan Huang
and
Liang Wang.
My research centers on Embodied Intelligence (Robotics & Game Agents) and Multimodal Models. My long-term goal is to develop generalist intelligent agents that assist people across both physical and virtual worlds, currently focusing on VLM as Policy and World Models.
2.1
Encoding linguistic semantics and visual signals within a shared latent space, enabling neural networks to comprehend the physical world across both spatial and temporal dimensions.
Leveraging upstream multimodal understanding to empower embodied agents—translating rich perceptual representations into world modeling, reasoning, and closed-loop action policies across physical robots and virtual environments.
3.1
No matching publications found.
4.1
We open-sourced NextTransformer, a compute-matched benchmark and autonomous research arena for next-generation Transformer architectures (Observatory).
2026.11Released two new preprints on embodied foundation agents: Scaling in Games and OneWM.
2026.10Video-Arena, our benchmark for spatiotemporal world state understanding and tracking, is now available on arXiv.
2026.10Concluded my research internship at Agibot Research.
2026.08EgoDemoGen is accepted to ACM MM 2026.
2026.07E-TTS is accepted to ECCV 2026.
2026.06BridgeVLA is accepted to NeurIPS 2025.
2025.115.1
Physical VLM Research, mentored by Shaoyu Chen.
Vision-Language-Action model pretrain.
Multimodal Agent research in the Visual Computing Group.
6.1