Abstract:
The computer-based real-time simulation of an electro-mechanical system using not only automatically fed pre-defined simulation conditions, but real life input signals to the system to enhance user control options is proposed. The introduced system is a robotic assistive technology for repair and maintenance of electrical power transmission lines in live condition. The setup is assisted by a simulation to provide real-time animation of the system under study while the controller operates the same, remotely and simultaneously. These systems can be used for research or educational purposes so as to measure the efficiency of the design for mechatronics industrial systems not only on predetermined inputs but on random test data sets as well and observe its response to the operational constraints that limit the system use.
Abstract:
The implementation and efficacy of closed loop systems in machines with moving limbs/parts is largely dependent upon the feedback systems that measure the extent of motion - linear or rotational. This paper proposes a novel technique for measurement of joint angles and thus
rotational motion for links pivoted at a powered/non-powered joint, using low-cost inertial sensors. The paper proposes the substitution of noisy, inefficient and poor resolution mechanical sensors like optical or pulse encoders with tri-axial accelerometers and tri-axial gyroscope fused in low-cost Inertial Measurement Units(IMUs). This technique is used to measure the angles between the various joints of a bipedal robot and estimate its complete orientation in 3D space. The crux of this paper is utilizing the extended capabilities of the inertial sensors in joint angle estimation for closed loop operation of a 12-Degree Of Freedom(DOF) lower body biped robot with potential implementation on stable bent knee walking on flat surfaces. All the joints of the biped are revolute and facilitate rotation of various limbs like thigh, shin and foot, analogous to a human leg. All links have IMUs mounted on them for the proposed task.
Abstract:
Engineering robot collaboration without significant reliance on sophisticated camera equipment to observe the robot's environment has always been a subject of interest for researchers. We believe combining proprioceptive sensing with reinforcement learning is essential for enabling robots to achieve tasks through collaboration in novel unknown environments. To that end, we propose a framework for two arm manipulators to work collaboratively in order to successfully transfer an object between locations that transcend one robot's reachability. Our major focus is on achieving mid-air object transfer, using a single RL agent, much like one brain controlling two arms. We attempt the proposed project in the V-REP simulation environment using deep Q-learning, a deep reinforcement learning algorithm and present the results of our experiments on subtasks of the whole task. Our agent successfully learned to place the object at the goal and to transfer the object among themselves to a reasonable degree.
