 |
 |
|
|
|
|
Current Plan: In the next fiscal year, we will fully realize the real-time reinforcement learning, using vision, on the real robot. Further, we will test the power of the two learning modes, imitation learning and reinforcement learning, using each mode separately and in combination. Within the next fiscal year, we will report the first case of real-robot vision online learning in which both learning modes are interleaved in real time: the imitation learning and reinforcement learning. We expect that the robot is able to run for hours to explore, to learn, and to practice, using powerful yet very challenging multiple sensors. Thanks to the task-nonspecific nature of our developmental program, in the next fiscal year we plan to demonstrate a variety of robot cognitive and behavioral capabilities, such as recognizing objects using vision, recognizing auditory commands using audition, performing autonomous navigation using vision and robot arm, detecting moving objects using vision, manipulation of recognized objects using vision and robot arm, and recognizing text inputs. All these capabilities will be learned in unknown environment. We will conduct rigorous performance measurement to gain knowledge about the quantitative relationships between a certain level of performance and the requirement for computational resources. For the deliberative layer, we intend to use the HHMM framework to learn and use a large multi-scale hierarchical map of the engineering building at MSU. This building is extremely large, with corridors that run for up to 400 feet, and would be extremely difficult to map using conventional "flat" models. We will also implement a hierarchical planning algorithm that will allow the robot to reach any selected destination in this environment starting from a state of no knowledge of its original starting location. We will use the HSM framework to learn a complex task of finding objects, picking them up, and depositing them into trash receptacles. Tasks of this complexity have never been learned before, because of the number of behaviors involved, and the problem of perceptual aliasing (e.g. if the robot sees a soda can, and turns around to avoid an obstacle, it needs to remember that it just saw a can). The trash collection task involves finding objects in the world through a pan-tilt-zoom camera, being able to move through the environment to approach an object that has been detected, picking it up, and finally depositing it into a trash collector. In the next fiscal year, the servo control of the autonomous mobile manipulators will be fully implemented. Making use of the redundancy of the mobile manipulator, hybrid force/position control will be achieved for moving objects. The coordination of the mobility and manipulability of the mobile manipulator will be tested also. The human/mobile manipulator action programming will be conducted such that the human action can be conveyed to the mobile manipulator via the Internet and the action will be executed autonomously. High level commands issued by the deliberative and reactive layers as well as the commands from human being will be integrated and transferred from the deliberative and the reactive layers to the servo control layer for execution. The feedback from lower layer will be conveyed to higher layer for interlayer interaction. The cooperation of multiple mobile manipulators/mobile robots will also be conducted for a variety of robotic tasks. In the 2nd year we will implement a realistic testbed for studying the SRS framework using complex state-of-the-art tasks. |
|