arxiv:2509.02530

Manipulation as in Simulation: Enabling Accurate Geometry Perception in Robots

Published on Sep 2

· Submitted by

ericonaldo on Sep 4

Upvote

Authors:

Minghuan Liu ,

Zhengbang Zhu ,

Xiaoshen Han ,

Jingxiao Chen ,

Yunfeng Lin ,

Xinghang Li ,

Bingyi Kang

Abstract

Camera Depth Models (CDMs) enhance depth camera accuracy by denoising and improving metric depth prediction, enabling better generalization of robotic manipulation policies from simulation to real-world tasks.

AI-generated summary

Modern robotic manipulation primarily relies on visual observations in a 2D color space for skill learning but suffers from poor generalization. In contrast, humans, living in a 3D world, depend more on physical properties-such as distance, size, and shape-than on texture when interacting with objects. Since such 3D geometric information can be acquired from widely available depth cameras, it appears feasible to endow robots with similar perceptual capabilities. Our pilot study found that using depth cameras for manipulation is challenging, primarily due to their limited accuracy and susceptibility to various types of noise. In this work, we propose Camera Depth Models (CDMs) as a simple plugin on daily-use depth cameras, which take RGB images and raw depth signals as input and output denoised, accurate metric depth. To achieve this, we develop a neural data engine that generates high-quality paired data from simulation by modeling a depth camera's noise pattern. Our results show that CDMs achieve nearly simulation-level accuracy in depth prediction, effectively bridging the sim-to-real gap for manipulation tasks. Notably, our experiments demonstrate, for the first time, that a policy trained on raw simulated depth, without the need for adding noise or real-world fine-tuning, generalizes seamlessly to real-world robots on two challenging long-horizon tasks involving articulated, reflective, and slender objects, with little to no performance degradation. We hope our findings will inspire future research in utilizing simulation data and 3D information in general robot policies.

View arXiv page View PDF Add to collection

Community

ericonaldo

Paper author Paper submitter 3 days ago

🚀 Want to build a 3D-aware manipulation policy, but troubled by the noisy depth perception? Want to train your manipulation policy in simulation, but tired of bridging the sim2real gap by degenerating geometric perception, like adding noise? Now these notorious problems are gone with our Camera depth Models! The Camera Depth Models (CDMs) can be plug-in modules in a real robot pipeline, transforming noisy depth into high-quality perception, enabling seamless sim-to-real transfer, making real robot manipulation work as in simulation!

🎯 Why it matters: Accurate geometry with CDMs helps a sim-data-driven policy solve a set of complex, long-horizon tasks from 0% to 85%+ success! Now you can even train in simulation, deploy on real robots WITHOUT further domain adaptation. Just plug in our CDMs to your existing pipeline!

✨ Highlight:
• Zero-shot sim-to-real transfer with 73%+ success (vs 0% baseline)
• Depth-only imitation learning achieves 85%+ success
• Works with RealSense D435/L515, Kinect, ZED2i & more

🛠️ Everything is open:
• Open-source CDMs for 5 distinct cameras
• Open-source the collected ByteCamDepth Dataset, which contains 170K+ RGB-depth pairs across 7 cameras & 10 configurations, a comprehensive real-world depth dataset.
• Open-source our codes for sim-to-real, camera depth model inference. We also share our modular, real-robot control framework designed for robotic manipulation, which provides you a unified interface for controlling various robot arms, integrating sensors, and executing policies in real-time!
• A clean sim-to-real tutorial based on our framework!

Check everything and interactive demos in https://manipulation-as-in-simulation.github.io/
We expect CDMs to be a foundation of your daily research!