Continual Learning Projects
ARNI Continual Learning Working Group: A Day in the Life
PI: Richard Zemel
Co-PI: Kathy McKeown
Abstract
Modern multimodal vision-language models (VLMs) are increasingly becoming capable of performing tasks across the spectrum of human interests, including games (e.g., card games, chess), abstract reasoning tasks (e.g., summarization, mathematical problem solving), and embodied intelligence tasks (pick-and-place). However, each of these generally requires a specialized VLM with task-specific training to perform well, and VLM post-training generally involves a single phase of fine-tuning/RL to achieve proficiency on one or a few narrow task domains. In contrast, humans simultaneously excel at many of these tasks, especially ones that are performed (and thus improved upon) frequently. As an example, the figure below (from here) shows that the average adult human spends several hours a day on household chores, work, education, and leisure, often interleaving these activities.
Publications
Continual Learning for Adaptive Human–Exosuit Control
PI: Nafiseh Ebrahami
Co-PI: Mohammadreza Davoodi and Xaq Pitkow
Abstract
Continual learning offers a principled approach for enabling adaptive systems to learn from ongoing experience while retaining previously acquired knowledge, a capability that is central to intelligent behavior in natural systems. However, its application to the control of safety-critical, human-in-the-loop robotic systems remains limited. Assistive exoskeletons, particularly soft wearable exosuits, provide a compelling and practically relevant testbed for advancing continual learning in such settings. Soft exoskeletons, or exosuits, offer a promising means to enhance human motor performance, reduce physical workload, and support rehabilitation by employing compliant materials that conform to the human body, preserve natural motion, and improve comfort and usability [1–4]. A fundamental challenge in exosuit control is that the interaction dynamics between the human and the exosuit are inherently nonstationary and continuously evolving. Changes in task objectives, external loads, posture, and fatigue can significantly alter the effective behavior of the coupled human–device system. These changes may occur gradually or abruptly and often without clearly defined task boundaries, making it difficult to predefine operating conditions or rely on fixed control strategies. As a result, controllers optimized for a single task or operating condition may exhibit degraded performance or unsafe behavior when deployed across varying conditions [5,6]. This motivates the need for adaptive control strategies that can learn continuously while retaining previously effective behaviors. Recent work has demonstrated the potential of continual learning for adaptive control of soft robotic systems, for example, by enabling neural-network-based controllers to adapt to changing external loads without catastrophic forgetting [7]. However, the application of continual learning to human–exosuit interaction, where learning must operate under physical interaction, user variability, and safety constraints, remains largely unexplored. To address this gap, this project aims to develop a unified framework for continual learning in human–robot interaction. An integrated soft wearable exosuit and musculoskeletal human model will serve as a human-in-the-loop evaluation platform. Specifically, the project leverages OpenSim-based musculoskeletal simulation [8] to model upper-body human dynamics coupled with a wearable exosuit and formulates exosuit assistance as a continual learning problem. OpenSim has been extensively used in prior studies to model assistive devices and to investigate their interaction with the human musculoskeletal system [9,10].
Publications
Curriculum Learning
PI: Richard Zemel
Co-PI: Natalie Brito
Abstract
We aim to develop a two-phase approach to model-specific curriculum learning, analogous to the process of scaffolding in child-caregiver interactions during infant development. In the discovery phase, we estimate the performance of a learning model on a task of interest based on its past performance and via targeted probing (e.g., question-answering). In the development phase, we leverage the information learned from the discovery phase on the performance of a
‘classroom’ of learning models and construct a curriculum that provides the maximum expected improvement in task performance for the learning models of interest. By iteratively applying this two-phase approach to a classroom of learning models, we aim to not only maximize performance and learning efficiency, but also to study the analogues and differences between humans and AI models during the learning process.
