Abstract-High precision assembly of mechanical parts requires accuracy exceeding the robot precision. Conventional part mating methods used in the current manufacturing requires tedious tuning of numerous parameters before deployment. We show how the robot can successfully perform a tight clearance peg-in-hole task through training a recurrent neural network with reinforcement learning. In addition to saving the manual effort, the proposed technique also shows robustness against position and angle errors for the peg-in-hole task. The neural network learns to take the optimal action by observing the robot sensors to estimate the system state. The advantages of our proposed method is validated experimentally on a 7-axis articulated robot arm.

While deep reinforcement learning techniques have recently produced considerable achievements on many decision-making problems, their use in robotics has largely been limited to simulated worlds or restricted motions, since unconstrained trial-and-error interactions in the real world can have undesirable consequences for the robot or its environment. To overcome such limitations, we propose a novel reinforcement learning architecture, OptLayer, that takes as inputs possibly unsafe actions predicted by a neural network and outputs the closest actions that satisfy chosen constraints. While learning control policies often requires carefully crafted rewards and penalties while exploring the range of possible actions, OptLayer ensures that only safe actions are actually executed and unsafe predictions are penalized during training. We demonstrate the effectiveness of our approach on robot reaching tasks, both simulated and in the real world.

No abstract

Physical risk factors assessment is usually conducted by analysing postures and forces implemented by the operator during a work-task performance. A basic analysis can rely on questionnaires and video analysis, but more accurate comprehensive analysis generally requires complex expensive instrumentation, which may hamper movement task performance. In recent years, it has become possible to study the ergonomic aspects of a workstation from the initial design process, by using digital human model (DHM) software packages such as Pro/ENGINEER Manikin, JACK, RAMSIS or CATIA-DELMIA Human. However, a number of limitations concerning the use of DHM have been identified, for example biomechanical approximations, static calculation, description of the probable future situation or statistical data on human performance characteristics. Furthermore, the most common DHM used in the design process are controlled through inverse kinematic techniques, which may not be suitable for all situations to be simulated. A dynamic DHM automatically controlled in force and acceleration would therefore be an important contribution to analysing ergonomic aspects, especially when it comes to movement, applied forces and joint torques evaluation. Such a DHM would fill the gap between measurements made on the operator performing the task and simulations made using a static DHM. In this paper, we introduce the principles of a new autonomous dynamic DHM, then describe an application and validation case based on an industrial assembly task adapted and implemented in the laboratory. An ergonomic assessment of both the real task and the simulation was conducted based on analysing the operator/manikin's joint angles and applied force in accordance with machinery safety standards (Standard NF EN ISO 1005-1 to 5 and OCcupational Repetitive Actions (OCRA) index). Given minimum description parameters of the task and subject, our DHM provides a simulation whose ergonomic assessment agrees with experimental evaluation.