University of Michigan Robotics · CMU Robotics · CMU ECE

📄 Preprint | 🎥 Video Demo

Check out the full instructions here 🚀🔥

Modern generative models like Sora produce stunning videos — but they often fall short in simulating the physics and dynamics of the real world.
This project highlights the strengths of physics-aware simulation via a customizable perturbation synthesis pipeline, enabling transformation from a Clean World to a Noisy World in a structured and controllable way.

Our pipeline synthesizes noisy RGB-D data for evaluating SLAM robustness under real-world perturbations:

1. Robot System & Trajectory Input → Global trajectory & system parameters.
2. Local Trajectory Generator → Physics engine simulates each sensor's motion.
3. Trajectory Perturbation Composer → Injects motion deviations.
4. Rendering Engine → Combines 3D scenes and perturbed sensor paths to produce clean sensor streams.
5. Sensor Perturbation Composer → Adds realistic corruptions to RGB-D streams.
6. Output → Fully perturbed datasets for robust SLAM benchmarking.

Robust SLAM is essential for real-world robot deployment. In this work:

• We introduce a modular pipeline to synthesize perturbations and evaluate SLAM robustness.
• A rich perturbation taxonomy and toolbox enable transformation from ideal simulations to challenging environments.
• We benchmark several top-performing RGB-D SLAM models under diverse, composed perturbations.
• Our analysis reveals model-specific vulnerabilities, often hidden under standard benchmark results.

• Perturbation Types: Evaluate under mixed or novel perturbations.
• Realism: Enhance the fidelity of simulated environments and distortions.
• Modalities: Extend beyond RGB-D — include LiDAR, sonar, etc.
• Model Development: Design more robust SLAM architectures.
• Broader Applications: Apply the evaluation to 3D reconstruction or navigation.

📌 See our paper for more details!

If you find Biblex helpful, please cite us:

@inproceedings{xu2025scalable,
  title     = {Scalable Benchmarking and Robust Learning for Noise-Free Ego-Motion and 3D Reconstruction from Noisy Video},
  author    = {Xiaohao Xu and Tianyi Zhang and Shibo Zhao and Xiang Li and Sibo Wang and Yongqi Chen and Ye Li and Bhiksha Raj and Matthew Johnson-Roberson and Sebastian Scherer and Xiaonan Huang},
  booktitle = {The Thirteenth International Conference on Learning Representations (ICLR)},
  year      = {2025},
  url       = {https://openreview.net/forum?id=Pz9zFea4MQ}
}

Got questions? Reach out to: xiaohaox@umich.edu

We gratefully acknowledge the following open-source projects:

• Classification-Robustness · Apache 2.0
• Replica Dataset · Research-only License
• Nice-SLAM · Apache 2.0
• Co-SLAM · Apache 2.0
• SplaTAM · BSD 3-Clause
• GO-SLAM · Apache 2.0
• ORB-SLAM3 · GPL v3

This project is licensed under the Apache License 2.0.