Uploaded May 2024 | Updated September 2026, 32 minutes ago
Microsoft Future Leaders in Robotics and AI Seminar Series: Towards Transparent & Reliable Embodied Reinforcement Learning Agents
Online Seminar
Thomas Berrueta
PhD Candidate
Northwestern University
The task-capabilities of robotic agents depend crucially on how their bodies materialize any action or decision made. Despite the clear importance of robot embodiment to decision-making, most algorithmic frameworks neglect or dismiss its impact. An important example of this lies in the widely used assumption that data are independent and identically distributed, which underpins all of machine learning. However, when data are collected sequentially from the experiences of agents this assumption does not generally hold, as is the case in most robotic reinforcement learning. In this talk, I will present a framework capable of overcoming these limitations by leveraging the statistical mechanics of ergodic processes, which I refer to as Maximum Diffusion Reinforcement Learning (MaxDiff RL). By measuring and exploiting statistical correlations, MaxDiff RL accounts for agent embodiment during learning and achieves provably robust state-of-the-art performance. Moreover, I will show that MaxDiff RL also provably enables agents to learn in continuous deployments. Taken together, I will argue that these results highlight future opportunities to formalize the role of robot embodiment during learning and control, paving the way towards more transparent and reliable embodied autonomy.
For more information on the Maryland Robotics Center see:
https://robotics.umd.edu
Microsoft Future Leaders in Robotics and AI Seminar Series: Towards Transparent & Reliable Embodied Reinforcement Learning Agents
Online Seminar
Thomas Berrueta
PhD Candidate
Northwestern University
The task-capabilities of robotic agents depend crucially on how their bodies materialize any action or decision made. Despite the clear importance of robot embodiment to decision-making, most algorithmic frameworks neglect or dismiss its impact. An important example of this lies in the widely used assumption that data are independent and identically distributed, which underpins all of machine learning. However, when data are collected sequentially from the experiences of agents this assumption does not generally hold, as is the case in most robotic reinforcement learning. In this talk, I will present a framework capable of overcoming these limitations by leveraging the statistical mechanics of ergodic processes, which I refer to as Maximum Diffusion Reinforcement Learning (MaxDiff RL). By measuring and exploiting statistical correlations, MaxDiff RL accounts for agent embodiment during learning and achieves provably robust state-of-the-art performance. Moreover, I will show that MaxDiff RL also provably enables agents to learn in continuous deployments. Taken together, I will argue that these results highlight future opportunities to formalize the role of robot embodiment during learning and control, paving the way towards more transparent and reliable embodied autonomy.
For more information on the Maryland Robotics Center see:
https://robotics.umd.edu
