Mengchen Ma

Mengchen Ma

M.S. student in Robotics Engineering @ Southeast University. Embodied AI & Reinforcement Learning.

Exploratory, Communicative, and Deployable: Vision-Driven Embodied Agents for Open-World Mobile Manipulation

Published in European Conference on Computer Vision (ECCV), 2026

REAL is a sim-to-real-consistent framework for interactive open-world mobile manipulation. Agents explore from raw RGB observations, use deployable navigation and manipulation tools through an MCP-based interface, and communicate with a simulated user to resolve ambiguous instructions — all without access to privileged simulator state.

The work contributes a hierarchical SFT and online RL training pipeline, and REAL-Bench, a benchmark of 241 tasks spanning four task families. The resulting agents reach 56.9% success on interactive tasks and 78.3% success across 60 real-world robot episodes.

Project Page · Code

Recommended citation: Boyu Mi*, Mengchen Ma*, Yifei Yao*, Xing Gao, Junting Chen, Yangzi Li, Zihou Zhu, Guohao Li, Zhenfei Yin, Tai Wang, Yao Mu, Jiangmiao Pang, Hanqing Wang. (2026). "Exploratory, Communicative, and Deployable: Vision-Driven Embodied Agents for Open-World Mobile Manipulation." ECCV 2026. (* Equal contribution)
Download Paper