Developing robust diffusion model for RGB based robot pose estimation
Published:
Abstract
This project focuses on improving PoseDiff - a diffusion-based model for estimating robot pose (6D end-effector pose and joint angles) using only RGB data from a fixed camera. PoseDiff achieves SOTA real-time performance using a DDPM based diffusion model conditioned with the RGB image embedding. There is potential to try various techniques used in typical diffusion models to explore its impact on PoseDiff’s performance. The ultimate goal is to create a more robust and efficient model capable of accurate RGB-only robot pose estimation assuming sensorless mobile manipulators.