Pranjali Pramod Pathre supervised by Dr. K Madhava Krishna  received her B.Tech (ECD). Here’s a summary of her research work on Streamlining Warehouse Operations: Monocular Multi-View Layout Estimation and Intelligent Visual Servoing for Robotic Tasks:

In today’s rapidly advancing technological landscape, where efficiency is paramount, there’s a pressing need for ongoing exploration into warehouse automation. As businesses worldwide contend with increasing consumer demands and the intricacies of supply chain management, it becomes crucial to devise intelligent strategies for optimizing warehouse operations. The momentum toward warehouse automation is accelerating, with projections indicating that robots could soon manage entire warehouse operations with minimal human intervention. However, it’s worth noting that a considerable proportion of warehouses operate without fundamental Warehouse Management Systems (WMS) in place. The issue of warehouse automation carries equal significance both in warehouses lacking WMS and in situations where automated robotic agents interact with shelf spaces. Performing actions like retrieving an item from a shelf, given the language command, can be significantly expedited through automated systems. Before engaging with an item on the shelf, the initial step involves reaching the object from within the warehouse. When provided with language instruction, the robot is expected to autonomously navigate and reach the desired target without human intervention. Motivated by this need, we introduce Imagine2Servo, an intelligent visual servoing-based controller designed to guide robots in reaching specific stacks of objects on shelves. This model isn’t solely intended for warehouse automation; it’s also tailored for various manipulation and long-range navigation tasks. Another key motivation behind the development of the Imagine2Servo model is to offer a significant enhancement to Image-based Visual Servoing (IBVS) algorithms. Traditional visual servoing algorithms rely on having a predefined goal image during testing. In Imagine2Servo, we innovate by generating a task-specific sub-goal image, which is then utilized by an IBVS controller to perform actions aimed at achieving the desired objective. While an autonomous agent operates within a warehouse space, it is also crucial to complement its actions with a semantic comprehension of the rack layout. This understanding, particularly regarding objects positioned on shelves, aids in addressing subsequent tasks such as determining object counts and estimating available space. To tackle this challenge, layout prediction comes into play, entailing the segmentation of all shelves within a rack at varying heights. While RackLay [1] provides layout estimation for racks, its scope is limited to predicting the layout of the predominant rack visible in the image. While suitable for constrained scenarios, this method inherently lacks scalability and adaptability, especially in dynamic warehouse settings where multiple racks may be concurrently present and only partially visible across a sequence of images. To tackle these challenges comprehensively, in we introduce MVRackLay, a multi-view layout estimation for all partly or wholly visible racks in the image. Furthermore, we leverage past observations to integrate spatial data over time, enhancing the accuracy of layout estimation. Additionally, we demonstrate the multi-view stitching of 3D layouts, which produces a 3D representation of the warehouse scene aligned with a global reference frame. 

 June 2024