IPAB Workshop - 06/10/2022

Speaker: Marina Aoyama

Title: "Efficient learning of adaptive behaviour involving force interaction"

Abstract: Today, robots are used in various aspects of our lives and there are many situations where humans and robots collaborate and work closely together. However, many challenges remain to be overcome before robots can be used in a wider range of fields. Robots that perform everyday tasks such as cooking, or that interact with humans in rehabilitation or assistive tasks, still have difficulty in identifying the characteristics of the manipulated object or human partner, that cannot be observed by vision alone, and adapting their behaviour accordingly. Therefore, my project aims to develop a method for robots to efficiently learn motor skills involving force interaction and adapt their behaviour according to characteristics of the manipulated object, such as its hardness or adhesiveness, through Learning from Demonstrations. 

Speaker: Ruchika Chavhan

Title: Dynamic Invariance Learning

Abstract: Standard paradigms for end-to-end supervised learning are notoriously data-hungry. Unfortunately, many application domains do not have access to big datasets because acquiring data involves a process that is expensive or time-consuming. This has motivated great interest in learning rich representations that can be transferred or reused across tasks. One of the factors that determines the usefulness of representations learned during pre-training for a given downstream task is the particular set of symmetries and invariances that are imbued within the model by pre-training. Providing invariances in a given learning task conveys a key inductive bias that can lead to sample-efficient learning and good generalization, if correctly specified. However, the ideal invariances for many problems of interest are often not known, which has led both to a large body of work to learn or incorporate desired invariances in models.  However, invariance learning is expensive and data intensive for popular neural architectures. A major vision in computer vision is that of general-purpose models: which can be cost-effectively used for a variety of diverse tasks. However, from the perspective of invariances and symmetries, I believe general-purpose models are insufficient for transfer on a diverse set of tasks. In this regard, I will talk about Efficient Dynamic Invariance Learning. I will discuss the limitations in some computer vision learning paradigms from the viewpoint of invariances and talk about some of my recent work on dynamic invariance learning. 

Speaker: Han Li

Title: Leveraging Task-Independent Information by Empowerment

Abstract: Information about the environment can be more significant to a robot than the reward in reinforcement learning in some cases. Empowerment is an important metric that represents the intrinsic information of the task environment with the reliability of the agent’s policy. It is independent of external rewards and presents a prominent example of intrinsic motivation. In this talk, I will discuss the benefits of Empowerment, demonstrate how it performs in simple environments, and consider extensions to higher-dimensional tasks.