Qihang Peng | 彭启航

I'm a senior undergraduate student at Xingjian College, Tsinghua University, and an incoming Ph.D. student at MMLab, CUHK.

My recent research mainly focuses on native multimodal pretraining and spatial intelligence. I also work with collaborators on visual generation, embodied intelligence, and LLM/VLM reasoning.

Now I'm working closely with Prof. Gao Huang and Prof. Hongsheng Li. If you are also interested in my research, please feel free to contact me.

CV  /  Google Scholar  /  Github  /  Twitter

WeChat: qihang_peng    Email: pengqihang22@gmail.com

Selected Publications

*Equal contribution

A generalist robotic manipulation foundation model that aligns representations, motions, and behaviors across 15 robot platforms. Trained on ~38,100 hours of pre-training data. Ranks first on RoboChallenge and surpasses π0.5 across 6 OOD benchmarks.

A unified navigation model built on Qwen3-VL spanning 5 task families with 15.6M training samples. Achieves SOTA on VLN-CE RxR, HM3Dv2, EVT-Bench, and NAVSIM with real-world deployment on Unitree Go2 quadruped robot.

A language-conditioned video world model with dual-stream MMDiT and MLLM action encoding, trained on 8.6M video-text pairs across 20+ robot morphologies. Ranks #1 on EWMBench, DreamGen Bench, and leads open-source models on WorldModelBench and PBench.

ColaVLA moves VLM reasoning into a compact latent space and decodes multi-scale causal trajectories in one pass. State-of-the-art in both open-loop and closed-loop settings with favorable efficiency and robustness on the nuScenes benchmark.

Make full use of multimodal information in ego-centric 3D visual grounding for point enhancement. State-of-the-art on the EmbodiedScan benchmark.

Use LLM and Ground Truth to enhance semantic details in prompt to reduce the ambiguity during training. Extract individual view semantics and enrich visual representation with global scene-level semantic.

Website Template