Tesi etd-01302026-120547
Link copiato negli appunti
Tipo di tesi
Corso di Dottorato (D.M.226/2021)
Autore
KUSHAWAHA, NILAY
URN
etd-01302026-120547
Titolo
Bridging Continual Learning and Robotic Intelligence: From Perception to Control
Settore scientifico disciplinare
ING-IND/34
Corso di studi
Istituto di Biorobotica - Ph.D. in Biorobotica
Relatori
relatore Prof. FALOTICO, EGIDIO
Parole chiave
- Continual Learning
- Reinforcement Learning
- Multimodality
- Data driven control
- Robotics
Data inizio appello
14/05/2026;
Disponibilità
parziale
Riassunto analitico
Most conventional machine learning algorithms assume access to data drawn from a stationary distribution, training a model once and retraining from scratch whenever the distribution shifts. Humans, by contrast, learn continuously and incrementally throughout life. Despite a gradual decline in learning efficiency with age, we rarely need to relearn previously acquired skills when encountering new ones. This remarkable ability to accumulate, refine, and integrate knowledge over time motivates the field of continual learning. The overarching goal of this thesis is to develop learning systems that emulate this lifelong learning capability, enabling artificial agents to acquire new skills without sacrificing existing competencies.
This dissertation aims to advance the state of the art in both perception and control by designing methods inspired by continual learning principles. We begin by establishing the foundations of continual learning, detailing its core problem settings, variants, and its relevance to robotics and perception. A central challenge—catastrophic forgetting, the degradation of performance on previously learned tasks when new tasks are acquired is examined in depth, along with prominent strategies proposed to mitigate it. We also provide a concise introduction to model-free, data-driven adaptive control approaches
relevant to robotic systems. The thesis contributions are organized into two major parts. The first focuses on enhancing robotic perception through brain-inspired continual learning algorithms capable of incremental object recognition from proprioceptive data. We further extend these capabilities by developing a multimodal continual learning framework that fuses
heterogeneous sensory modalities sequentially, eliminating the need for full retraining when incorporating new sources of information.
The second part develops adaptive continual learning algorithms for robot control. We introduce a dynamic policy capable of learning multiple manipulation tasks sequentially in simulation via reinforcement learning and efficiently adapting to real-world tasks through imitation learning, enabled by strong sim-to-real knowledge transfer. Additionally, we design a dynamic continual control framework for modular soft robots, in which each module is governed by a dedicated sub-network. This distributed architecture enhances precision, flexibility, and scalability compared to monolithic controllers. Across perception and control, this thesis demonstrates how continual learning provides a powerful and biologically inspired paradigm for building adaptive, scalable, and memory-efficient robotic systems. The presented methods highlight the potential of continual learning to fundamentally reshape how artificial agents learn, generalize, and interact with the world.
This dissertation aims to advance the state of the art in both perception and control by designing methods inspired by continual learning principles. We begin by establishing the foundations of continual learning, detailing its core problem settings, variants, and its relevance to robotics and perception. A central challenge—catastrophic forgetting, the degradation of performance on previously learned tasks when new tasks are acquired is examined in depth, along with prominent strategies proposed to mitigate it. We also provide a concise introduction to model-free, data-driven adaptive control approaches
relevant to robotic systems. The thesis contributions are organized into two major parts. The first focuses on enhancing robotic perception through brain-inspired continual learning algorithms capable of incremental object recognition from proprioceptive data. We further extend these capabilities by developing a multimodal continual learning framework that fuses
heterogeneous sensory modalities sequentially, eliminating the need for full retraining when incorporating new sources of information.
The second part develops adaptive continual learning algorithms for robot control. We introduce a dynamic policy capable of learning multiple manipulation tasks sequentially in simulation via reinforcement learning and efficiently adapting to real-world tasks through imitation learning, enabled by strong sim-to-real knowledge transfer. Additionally, we design a dynamic continual control framework for modular soft robots, in which each module is governed by a dedicated sub-network. This distributed architecture enhances precision, flexibility, and scalability compared to monolithic controllers. Across perception and control, this thesis demonstrates how continual learning provides a powerful and biologically inspired paradigm for building adaptive, scalable, and memory-efficient robotic systems. The presented methods highlight the potential of continual learning to fundamentally reshape how artificial agents learn, generalize, and interact with the world.
File
| Nome file | Dimensione |
|---|---|
Ci sono 1 file riservati su richiesta dell'autore. |
|