Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Continual Learning for Autonomous Robots: A Prototype-based Approach

Elvin Hajizada, Balachandran Swaminathan, Yulia Sandamirskaya

TL;DR

The paper addresses open-world, few-shot online continual learning for autonomous robots by introducing Continually Learning Prototypes (CLP), a rehearsal-free, prototype-based approach augmented with per-prototype metaplasticity to balance plasticity and stability. CLP supports on-demand multi-prototype per class, novelty detection, and semi-supervised learning, with a framework compatible with neuromorphic hardware. Key contributions include a dynamic learning-rate mechanism for prototypes, open-world novelty handling, and empirical validation on OpenLORIS showing state-of-the-art performance for FS-OCL and strong novelty-detection metrics, plus a neuromorphic Lava/Loihi 2 implementation. The work demonstrates practical impact by enabling robust, real-time lifelong learning for robots, with promising directions for integrating object detection and fully neuromorphic deployment.

Abstract

Humans and animals learn throughout their lives from limited amounts of sensed data, both with and without supervision. Autonomous, intelligent robots of the future are often expected to do the same. The existing continual learning (CL) methods are usually not directly applicable to robotic settings: they typically require buffering and a balanced replay of training data. A few-shot online continual learning (FS-OCL) setting has been proposed to address more realistic scenarios where robots must learn from a non-repeated sparse data stream. To enable truly autonomous life-long learning, an additional challenge of detecting novelties and learning new items without supervision needs to be addressed. We address this challenge with our new prototype-based approach called Continually Learning Prototypes (CLP). In addition to being capable of FS-OCL learning, CLP also detects novel objects and learns them without supervision. To mitigate forgetting, CLP utilizes a novel metaplasticity mechanism that adapts the learning rate individually per prototype. CLP is rehearsal-free, hence does not require a memory buffer, and is compatible with neuromorphic hardware, characterized by ultra-low power consumption, real-time processing abilities, and on-chip learning. Indeed, we have open-sourced a simple version of CLP in the neuromorphic software framework Lava, targetting Intel's neuromorphic chip Loihi 2. We evaluate CLP on a robotic vision dataset, OpenLORIS. In a low-instance FS-OCL scenario, CLP shows state-of-the-art results. In the open world, CLP detects novelties with superior precision and recall and learns features of the detected novel classes without supervision, achieving a strong baseline of 99% base class and 65%/76% (5-shot/10-shot) novel class accuracy.

Continual Learning for Autonomous Robots: A Prototype-based Approach

TL;DR

The paper addresses open-world, few-shot online continual learning for autonomous robots by introducing Continually Learning Prototypes (CLP), a rehearsal-free, prototype-based approach augmented with per-prototype metaplasticity to balance plasticity and stability. CLP supports on-demand multi-prototype per class, novelty detection, and semi-supervised learning, with a framework compatible with neuromorphic hardware. Key contributions include a dynamic learning-rate mechanism for prototypes, open-world novelty handling, and empirical validation on OpenLORIS showing state-of-the-art performance for FS-OCL and strong novelty-detection metrics, plus a neuromorphic Lava/Loihi 2 implementation. The work demonstrates practical impact by enabling robust, real-time lifelong learning for robots, with promising directions for integrating object detection and fully neuromorphic deployment.

Abstract

Humans and animals learn throughout their lives from limited amounts of sensed data, both with and without supervision. Autonomous, intelligent robots of the future are often expected to do the same. The existing continual learning (CL) methods are usually not directly applicable to robotic settings: they typically require buffering and a balanced replay of training data. A few-shot online continual learning (FS-OCL) setting has been proposed to address more realistic scenarios where robots must learn from a non-repeated sparse data stream. To enable truly autonomous life-long learning, an additional challenge of detecting novelties and learning new items without supervision needs to be addressed. We address this challenge with our new prototype-based approach called Continually Learning Prototypes (CLP). In addition to being capable of FS-OCL learning, CLP also detects novel objects and learns them without supervision. To mitigate forgetting, CLP utilizes a novel metaplasticity mechanism that adapts the learning rate individually per prototype. CLP is rehearsal-free, hence does not require a memory buffer, and is compatible with neuromorphic hardware, characterized by ultra-low power consumption, real-time processing abilities, and on-chip learning. Indeed, we have open-sourced a simple version of CLP in the neuromorphic software framework Lava, targetting Intel's neuromorphic chip Loihi 2. We evaluate CLP on a robotic vision dataset, OpenLORIS. In a low-instance FS-OCL scenario, CLP shows state-of-the-art results. In the open world, CLP detects novelties with superior precision and recall and learns features of the detected novel classes without supervision, achieving a strong baseline of 99% base class and 65%/76% (5-shot/10-shot) novel class accuracy.
Paper Structure (18 sections, 8 equations, 7 figures, 1 table)

This paper contains 18 sections, 8 equations, 7 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related work
  3. Problem Setting
  4. Continually Learning Prototypes
  5. Prototype-based Learning
  6. Online Learning with Prototypes
  7. Few-shot Learning with Prototypes
  8. Metaplasticity for Continual Learning
  9. Open-world Recognition and Semi-supervised Learning
  10. Learning Multi-modal Representations
  11. Experiments
  12. General Experimental Setting
  13. Fully Supervised Online Continual Learning
  14. Open-set Recognition with Novelty Detection
  15. Semi-supervised Few-shot Continual Learning
  16. ...and 3 more sections

Figures (7)

  • Figure 1: Continually Learning Prototypes: overview
  • Figure 2: Fully supervised online continual learning of all classes in the OpenLORIS dataset.
  • Figure 3: T-SNE visualization of the OpenLORIS features extracted by pre-trained EfficientNet-B0 backbone and learned prototypes. We randomly chose ten videos (60 frames each) for all 40 classes. The videos from each category may include different object instances but also variations of the same instances. As pointed out in the figure, some classes are represented with a single cluster and hence a single prototype (e.g., class 13), while others are clustered into varying numbers of clusters (e.g., class 4, 7, and 18) and accurately represented by multiple prototypes. This demonstrates that the methods that learn each class with a single prototype are inadequate. Conversely, CLP has adequate representational power thanks to its per-class, on-demand, multi-prototype learning mechanism.
  • Figure 4: Performance analysis of CLP's novelty detection mechanism for open-set recognition: (a) ROC curve, (b) precision-recall curve, (c) precision-recall curve with persistent threshold
  • Figure 5: Few-shot semi-supervised continual learning. CLP retains base class accuracy while learning novel classes without supervision. As the number of videos provided per novel class increases, the accuracy improves significantly.
  • ...and 2 more figures