← Back ICRA 2026

UniDoorManip: Learning Universal Door Manipulation Policy Over Large-Scale and Diverse Door Manipulation Environments

Yu Li, Xiaojie Zhang, Ruihai Wu, zilong zhang, Yiran Geng, Hao Dong, Zhaofeng He

PDF

AI summary

Key figure (auto-extracted from paper)

A three-stage disentanglement framework with conditional training enables robots to universally manipulate diverse, unseen doors across simulation and real-world settings.

Universal manipulation Door manipulation Visual affordance Conditional training Simulation-to-real Articulated objects

Problem

Prior datasets and simulations lack geometric and mechanistic diversity, preventing robots from learning universal policies that generalize to unseen doors and complex real-world scenarios.

Approach

The authors introduce a large-scale diverse door dataset and realistic simulation, then propose a framework that splits manipulation into grasping, handle manipulation, and door opening stages, integrating them through conditional training.

Key results

Large-scale dataset with 6 door categories and thousands of compositional variants
Realistic simulation with diverse latching mechanisms and partial occlusion
Strong generalization to unseen door geometries and categories in simulation
Successful real-world transfer of the universal manipulation policy

Why it matters

Provides a scalable foundation for embodied agents to reliably navigate and interact with the wide variety of doors found in complex real-world environments.

Abstract

Learning a universal manipulation policy encom- passing doors with diverse categories, geometries and mecha- nisms, is crucial for future embodied agents to effectively work in complex and broad real-world scenarios. Due to the limited datasets and unrealistic simulation environments, previous stud- ies fail to achieve good performance across various doors. In this work, we build a novel door manipulation environment reflecting different realistic door manipulation mechanisms, and further equip this environment with a large-scale door dataset covering 6 door categories with hundreds of door bodies and handles, making up thousands of different door instances. To learn a universal policy over diverse doors, we propose a novel framework disentangling the whole manipulation process into three stages, and integrating them through conditional training. Extensive experiments validate the effectiveness of our designs and demonstrate our framework’s strong performance in simulation and real world. Code, data and videos are available on https://unidoormanip.github.io/.

Index terms

Perception for Grasping and Manipulation