Dartmouth Events

Abstract: Bringing the power of AI from the virtual realm into the physical world holds transformative potential to redefine how we accomplish everyday tasks and deepen our understanding of intricate natural systems around us.  This talk will explore two dual notions of physical intelligence that aim to move towards this goal: (1) creating more natural robot assistants by learning from human demonstrations, and (2) using robots to study sperm whale physiology and social behavior.  In both cases, smart wearable or deployable sensing will be used to fill data gaps and support machine learning pipelines.

The first part of the talk will focus on designing adaptive robot assistants that perform household tasks with human-like movements to facilitate human-robot collaboration.  A multimodal dataset and open-source recording framework, dubbed ActionSense, curates human demonstrations of common kitchen tasks.  A recently developed machine learning architecture, a physics-inspired state-space model, is then applied to this data to teach robots from these demonstrations and generate trajectories that mimic key characteristics of the human motions.  

The second part of the talk will focus on using machines to study nature - it will describe how deployable sensing and machine learning pipelines can provide insights into the lives of sperm whales.  Project CETI is creating open-source wearable sensor tags and distributed robot systems that can scale data foundations to probe outstanding questions about whale physiology and behavior.  Sample results from applying machine learning to a rare observation of a sperm whale birth will be presented...

Bio: 

https://www.josephdelpreto.com/