Improving the Perception of Autonomous Robotic Systems through Sensing and Machine Learning

最近，机器人领域经历了一个戏剧性的激增，无论是在科学成就和工业应用方面。许多任务的自动化已经被提出，例如驾驶、对象操作或仓库操作。为了减轻与不确定性或不断变化的条件相关的困难，有必要改进感知管道。在这个提案中，我们将在智能层面和感知层面对其进行改进。对于前者，我们将探索两种机器学习技术的使用：领域适应（DA）和稀疏编码（SC）。领域自适应旨在提高在数据集上训练的分类器的性能，但在分布略有不同的数据上使用。稀疏编码试图通过寻找潜在长特征向量中很少成分活跃（非零）的表示来自动化特征提取问题。例如，在传感层面，我们建议探索使用定制的高光谱相机。

Recently, the field of robotics has experienced a dramatic surge, both in terms of scientific accomplishments and industrial applications. The automation of numerous tasks has been proposed, such as driving, object manipulation or warehouse operation. To mitigate the difficulties associated with uncertainties or changing conditions, improvements to the perception pipeline are necessary. In this proposal, we will improve it both at the intelligence level and at the sensing level. For the former, we will explore the use of two machine-learning techniques: Domain Adaptation (DA) and Sparse Coding (SC). Domain Adaptation aims at improving performance of a classifier trained on a data set but used on data which is distributed slightly differently. Sparse Coding tries to automate the problem of feature extraction, by finding representations where few components of a potentially long feature vector are active (non-zero). At the sensing level, we propose for example to explore the use of custom-made hyperspectral cameras. For our short term objectives, we have identified 3 key problems in robotics for which we seek to make significant contributions: visual place recognition, autonomous navigation in forests, and grasping automation. The use of DA in place recognition will improve the detection of images of the same location taken under different illumination or weather conditions. In order to remove shadows in images, better color-constant images can be generated from hyperspectral cameras. For forested environments, we will study, in parallel, the use of this hyperspectral camera and of 3D LiDAR scans for place recognition. On top of that, we will propose ameliorations to the creation process of topometric maps, used for forest navigation purposes. For grasping, we propose a richer representation of a grasping location, called hemicylindrical view. We will also increase the robustness of the 3D sensing by fusing multiple views. Finally, we will perform representation learning with Sparse Coding. Experiments will be conducted on real data or robots. For instance, we will regularly gather datasets in Quebec City, over the four seasons to test our visual place recognition methods. For forest navigation, we will use our Clearpath Robotics Husky A200 robot and its sensor suite in the forests located on Laval University campus. For grasping, we will test our algorithms with real robotic arms. We expect to make significant scientific contributions to robotics, in the form of novel applications of advanced machine learning methods or sensing approaches. For example, we do not believe that anyone has explored the paradigm of Domain Adaptation in the context of place recognition. We also expect that our research results will transfer directly to the industry. Finally, we will train 3 Undergraduates, 2 Masters and 4 PhDs with skills and knowledge that will benefit the Canadian industry.