ARIAM Seminar Series: Motion Planning in Constrained Environments

Abstract:

Diffusion Models are deep learning models that have been very successful in generative modelling of complex data such as images, videos and text. They have also shown state of the art performance in difficult robotics tasks such as imitation learning, but they discount prior knowledge of system dynamics and task objectives. Model-Based Diffusion is a technique that offers an alternative approach enabling diffusion based on models while retaining the ability to incorporate learning. However, Model-Based Diffusion struggles in severely constrained environments due to its Monte Carlo sampling-based mature. This presentation introduces diffusion models in the context of optimisation theory, how model-based diffusion can be used to do motion planning, why it struggles in constrained environments, and how it can be improved. We present applications and simulated examples for underwater robotic manipulation.

Bio:
Raghav is a second year PhD researcher in Control and Motion Planning at ARIAM Research Hub.  He received a Bachelors and Masters in Mechatronic Engineering at the University of Queensland in 2021. As part of the ARIAM Hub he is currently working with industry partner Reach Robotics on combining learning and planning for underwater mobile manipulation.