Principal Investigator:
Dr. Hasan Sajid
Department of Robotics and Artificial Intelligence, SMME
National University of Sciences and Technology (NUST)
Main Campus, Sector H-12, Islamabad.
Phone: +92-51-90856080, Email: hasan.sajid@smme.nust.edu.pk
Goal: Application of AI to solve fundamental problems in the field of robotics
Scientific Objectives:
- General policy learning for robotic tasks with focus on motion planning
- Generalization from less data
Description of Lab:
Intelligent robotics is widely believed to spearhead the upcoming technological revolution where robots and intelligent machines will become an integral part of everyday human life. Ranging from area specific domains such as industrial automation, manufacturing, entertainment, defense and surgical robotics to machines that operate in common everyday human households, robots can already be seen assisting humans in virtually every aspect of their lives. Not only do they help reduce human effort but also, at the same time, improve the precision and accuracy with which these tasks are accomplished. It is for these reasons that modern day researchers, industrialists, investors and analysts are all unequivocally predicting robotics as the next big thing and one of the very central domains of Artificial Intelligence (AI) for this new era.
The current robotics trend has been dominated by sense-plan-act paradigm, where robot observes the environment around it (perception/sensing), keeps track of its internal state (sensing), forms a plan of action (plan) and then executes the plan (act). Such an approach has been very effective for problems in controlled environments but fail to generalize to real world conditions. For example, scene understanding, a core task of perception/sensing module, works only in controlled environmental conditions, and breaks down in real world conditions due to challenges such as illumination changes, shadows, occlusions, diverse weather conditions and camera perspective to name a few. The generalization problem is equally applicable to planning module. Examples are basic object grasping task and advanced driverless cars, where the planning module has to map from very high dimensional sensory data to low dimensional joint space.
AI has lead researchers to solve some of the problems by building better and more powerful end-to-end models. For example in case of scene understanding, the recent AI algorithms have shown remarkable robustness in real world conditions [1] leading to its wide adaptability among roboticists. Likewise, there have been recent attempts to solve planning problems in the form of grasping [2] and motion planning [3]. However due to supervised nature of solutions, these require significantly large amounts of real world data, which is costly and even impossible in some cases. Examples include the ImageNet dataset [1] with over 14 million hand-annotated images for simple visual classification problem, the grasping dataset [4] comprising of 650K grasp attempts for robotic grasping and the driverless car datasets that require driving a fleet of cars rigged with costly sensors (LIDAR + radar + cameras + Others) millions of miles. Such data can only be generated with significant resources and time.
To overcome the problems associated with conventional robotic approach and current AI approaches, it is imperative to develop foundational algorithms and framework that can learn general policy for robotic tasks with less amount of data. The wide applicability of such algorithms ranging from variety of analytics to different types of robotic tasks will open up new boundaries for scientific research and industrial applications.
Applications Domains:
-
- Autonomous surveying
- Driverless vehicles
- Fault detection in manufacturing industries for automatic QA
- Crop Management and Estimation
- Warehouse management