Related Experiment Video
Updated: Aug 6, 2026
07:41
Design and Implementation of a Bespoke Robotic Manipulator for Extra-corporeal Ultrasound
Published on: January 7, 2019
Depth-augmented diffusion policy with pseudo-depth for robust robotic manipulation
Lesia Hentosh1, Yaroslav Hladun2, Ivan Tsmots2
1Lviv Polytechnic National University, Lviv, 79013, Ukraine. lesia.i.mochurad@lpnu.ua.
Scientific Reports
|July 16, 2026
Summary
Depth-augmented Diffusion Policy improves robotic manipulation by using depth data alongside RGB images. This framework effectively uses pseudo-depth, removing the need for RGB-D cameras and boosting task success rates.
Area of Science:
- Robotics
- Computer Vision
- Machine Learning
Background:
- Diffusion Policy is effective for robotic manipulation imitation learning.
- RGB-only policies struggle with tasks requiring precise spatial understanding and distance-aware motions.
- Integrating depth data presents challenges in conditioning mechanisms and depth reliability.
Purpose of the Study:
- To develop a Depth-Augmented Diffusion Policy framework for enhanced robotic manipulation.
- To enable effective utilization of depth information in diffusion policies.
- To investigate the use of pseudo-depth as an alternative to sensor depth.
Main Methods:
- Employs separate visual encoders for RGB and depth data.
- Fuses representations via stochastic feature-level addition with modality dropout.
- Utilizes pseudo-depth generated by Depth Anything V2 estimator.
Main Results:
- Achieved 95% success rate with sensor depth and 90% with pseudo-depth on a toilet-seat closing task.
- Outperformed RGB-only baseline (50%) and 3D Diffusion Policy (75%).
- Pseudo-depth provided consistent improvements on the LIBERO simulation benchmark.
Conclusions:
- Explicit depth conditioning significantly enhances performance on distance-sensitive robotic manipulation tasks.
- Pseudo-depth serves as a viable alternative to hardware depth, reducing reliance on specialized cameras.
- The proposed framework demonstrates broad applicability and effectiveness across various robotic tasks.
