Issue link: https://htpgraphics.uberflip.com/i/1385717
LEAD RESEARCHER: PRAGNA DAS I am a robotics researcher in mobile robots (MR), multi- robot system (MRS), multi-agent system (MAS), arm manipulators, specializing in physical and environmental cost parameter modelling and estimation, adaptive decision making based on cost parameter, collective Intelligence, mistake aware learning by demonstration, deep learning, path planning, collision avoidance, navigation for both arm manipulators and mobile robots and indoor logistics for smart factories. Experience in Pytorch, Scikitlearn in the current work related to collision avoidance, path and task planning. COLLISION AVOIDANCE IN CLUTTERED ENVIRONMENTS UNIQUENESS // The goal is to avoid collision of obstacles with the whole arm. Potential options such as sensitive skin need sophisticated hardware, whereas the potential field methods are computationally expensive. Also, the available inverse kinematics methods do not encompass collision detection with multiple-section (links) robotic arms. A key to the solution is using a cost of collision, based on the distance between the links and the objects. The cost function takes the current configuration and an object as input and provides the cost of collision. The uniqueness is training the neural network using data from a simulated glovebox to mimic its collision cost function. The cost of collision is obtained four times faster than the analytical cost function computation. This cost of collision forms the constraint for joint poses in the inverse kinematics optimisation. Additionally, an open-source simulated glovebox environment and a synthetic dataset for obstacle in glovebox has been created to foster further research. 52 SUMMARY // The numerous objects inside a glovebox cause obstruction towards the free movement of a manipulator to carry out POCO processes inside the glovebox. When the end- effector of the robotic arm needs to reach its goal pose, the arm must calculate the change of joint angles at each update. A set of achievable joint poses or motion trajectories should be determined, so that the end joint pose places the end-effector to the goal pose. With limited visibility, due to partial darkness and clutter, the propensity of collision in this case is along the whole body of the arm. The conventional approaches do not work either due to low visibility or lack of apriori knowledge of the environment. My work determines the cost of collision between the whole body of the manipulator and the clutter objects using a neural network cost approximation method for fast collision detection. A dataset for the training of neural network has been created and the method was tested in both real and simulated mock-up nuclear glovebox environments. Through this model, the cost of collision is obtained four times faster than the analytical cost function computation.