Imitation Learning for High Precision Peg-in-Hole Tasks

A Survey of Robot Manipulation in Contact

“…Finally, there are tasks which could be performed without any control of the forces under perfect knowledge, such as the classical peg-inhole and similar workpiece alignment tasks and articulated motions, such as opening a door. However, any uncertainty in such tasks raises the need for controlling the contact forces to prevent excessive collisions; moreover, by leveraging compliance a robot can perform, for example, a peg-in-hole task with clearance smaller than the robot's accuracy [17]. In this section more details of the manipulation skills requiring controlling of force interactions will be given under three categories; environment shaping, workpiece alignment and articulated motions.…”

“…Regardless of the long history, a human can still outperform a robot in peg-in-hole tasks in certain metrics, such as generalization, managing surprising situations and uncertainties and really tight clearances [44]; however, there is work to overcome these, such as meta-reinforcement learning for generalization [45] and clearances smaller than the robot's accuracy (6µM ) [17]. There are also more difficult variants, such as multi-peg-in-hole [42], assembly construction [122], hole-in-peg with threaded parts [123] or pegin-hole combined with articulated motions, which include as folding [75] or snapping [76,77] where either more elaborate motions or force over a certain threshold is required to complete the task.…”

On the Improvement of ROS-Based Control for Teleoperated Yaskawa Robots

“…Furthermore, the combination of system control with even small time delays creates stability problems [13], which has led to alternative control approaches. Robotic teleoperation is proposed in [14] for imitation learning data collection. Such method represents stochastic artificial intelligence approaches and is used for complex manipulation tasks.…”

Translating Natural Language Instructions to Computer Programs for Robot Manipulation

“…The majority of recent works on imitation learning have used some input device such as game controller [9], VR controller [3], visual odometry based 6-DoF position tracking using smartphones [10] [11], space mouse [12], etc. to record experts teleoperating a robot.…”

“…A few examples of the tasks we consider are: (a) Push the orange towards the apple, (b) Place the apple between the orange and the apple, (c) Pick up the orange and use it to push the bottle off the edge of the table. Although our robot does not use a force sensor and can only move the endeffector using position control, it is possible to expand the set of primitive instructions of the robot to include complex macro instructions such as peg-in-hole insert instruction that may invoke a separately trained policy network [12]. Our approach is most suitable for "gluing" together simpler commands to compose a more complex program.…”