EmbodMocap: In-the-Wild 4D Human-Scene Reconstruction
for Embodied Agents
(CVPR2026)

Wenjia Wang¹ Liang Pan¹ Huaijin Pi¹ Yuke Lou¹ Xuqian Ren² Yifan Wu¹
Zhouyingcheng Liao¹ Lei Yang³ Rishabh Dabral⁴ Christian Theobalt⁴ Taku Kumora^1†

¹The University of Hong Kong ²Tampere University ³Chinese University of Hong Kong ⁴Max-Planck-Institute for Informatics

arXiv(Coming Soon) PDF Code(Coming Soon) Data(Coming Soon)

Your Image

4D Human & Scene Reconstruction results:

Your Image

The dataset captured by our method:

Captured Seq
0505step2_seq0:

Captured Seq
0615stairs1_seq3:

Captured Seq
0618stair1_seq0:

Captured Seq
0902bedroom1_seq8:

Captured Seq
0914livingroom1_seq13:

Motion tracking results:

Scene Aware Tracking
0505step2_seq0:

Scene Aware Tracking
0615stairs1_seq3:

Scene Aware Tracking
0902bedroom1_seq8:

Scene Aware Tracking
0914livingroom1_seq13:

HOI Skill results:

Climb / monocular data:

Climb / opticalmocap data:

Climb / our data:

Sit / monocular data:

Sit / opticalmocap data:

Sit / our data:

Prone / monocular data:

Prone / our data:

Support / monocular data:

Support / our data:

Support2 / monocular data:

Support2 / our data:

Humanoid Robot results:

Seq 1 / Squat:

Seq 2 / Lunge:

Seq 3 / Carwheel: