Incremental robot learning of new objects with fixed update time

Abstract

We consider object recognition in the context of lifelong learning, where a robotic agent learns to discriminate between a growing number of object classes as it accumulates experience about the environment. We propose an incremental variant of the Regularized Least Squares for Classification (RLSC) algorithm, and exploit its structure to seamlessly add new classes to the learned model. The presented algorithm addresses the problem of having an unbalanced proportion of training examples per class, which occurs when new objects are presented to the system for the first time.We evaluate our algorithm on both a machine learning benchmark dataset and two challenging object recognition tasks in a robotic setting. Empirical evidence shows that our approach achieves comparable or higher classification performance than its batch counterpart when classes are unbalanced, while being significantly faster.

Interactive data collection for deep learning object detectors on humanoid robots

“…Finally, we plan to investigate possible strategies to speedup the current training algorithm in order to provide the robot with the ability of learning on the fly to detect more objects, while interacting with the human. A possible approach which we will evaluate is to start from recent work on incremental object recognition [5] and apply a similar approach to the object detection problem. This would allow to incrementally update the detector, rather than training it from scratch every time new image examples are collected by the robot.…”

“…We manually annotated this subset of images from the MIX sequences used in the previous test. We adopted the labelImg tool 5 and fixed an annotating policy such that an object must be annotated if at least a 50-25% of its total shape is visible (i.e. not cut out from the image or occluded).…”

AbdomenCT-1K: Is Abdominal Organ Segmentation a Solved Problem?

“…The learned knowledge of a model can be interferes when we train the model with new information, which can cause performance decrease for the old task. Therefore, continual learning has attracted growing attention in the past years [55], such as object recognition [56], [57] and classification [58]. Besides, tailored datasets and benchmarks for continual learning have been also proposed in computer vision community, e.g.…”

CRIL: Continual Robot Imitation Learning via Generative and Prediction Model

“…[9] and [22] use rehearsal to store all the training data of previous tasks in a buffer and then use them in the training of new tasks. [23] and [24] use state representation learning to build a general perception model for objects and environments, and then use the representation to facilitate the new task learning process, such as curiosity-driven exploration [25]. However, these methods have different assumptions that make them impractical and lack scalability in real-world robot tasks: having a large storage space [8], [9], [22], having access to previous environments [6], [7], [21], or similar skills can be adopted in different tasks [23], [24].…”