Learning admittance mappings for force-guided assembly

Abstract

We present a practical method for autonomous synthesis of appropriate admittance behavior for robust high-precision robotic assembly. Because our approach relies on on-line learning of the appropriate admittance through repeated attempts at the assembly operation, we are able to circumvent the problems alternative approaches have in trying to model the interactions between the robot and its environment. Test results on the peg-in-hole insertion task show that the performance of our approach compares favorably with that of other methods recently proposed for highprecision chamferless peg-in-hole insertion.A primary concern in robotic assembly is to overcome part misalignment that can cause the assembly operation to fail. Precision assembly operations are particularly susceptible to misalignment. As the clearance between parts decreases, part misalignments become increasingly critical because the precision required approaches the level of uncertainty in positioning the parts. In such cases, conventional position control of the assembly operation has to be augmented with some method of overcoming possible misalignments. One approach is to use geometric path planning techniques to determine a motion sequence that has the highest likelihood of success given the part geometries and the margins of uncertainty in executing that sequence. Pre-imaging [12], backprojection [B], and other strategies (e.g., [5, 71) have been proposed for geometric path planning.Alternative approaches, based on reactive control, try to counter the effects of uncertainty with on-line modification of the assembly path based on sensory feedback. Compliant motion control, in which the motion path is modified by contact forces or tactile stimuli occurring during the motion, is often used. In this paper, we focus on force-guided assembly strategies that cause the manipulator to deviate from its nominal motion path depending on the sensed contact forces.Force control serves two purposes. First, it can be used to prevent excessive contact forces from damaging the parts or the manipulator during assembly. More importantly, force control can correct part misalignment by mapping contact forces arising from misalignment to motions that reduce the misalignment.A key component in implementing force-guided assembly therefore is to specify the manipulator's admittance, which determines how the manipulator responds to forces. The specified admittance behavior can be realized in several ways. Passive mechanical devices such as the remote center compliance (RCC)[18] can be used to implement fixed diagonalized linear admittance mappings. When different admittances are to be used for assembling different parts, more elaborate active compliance devices can be programmed to exhibit the desired admittance characteristics. The admittance also can be implemented in software, which has the advantage that complex, nonlinear admittances can be realized without resorting to special hardware.The difficulty, however, lies in determining the appropriate admittance to use when...

Compliant manipulation for peg-in-hole: Is passive compliance a key to learn contact motion?

“…Considering the width of the peg(40mm) in the simulations, this noise is large. Note that the previous research on reinforcement learning deal with much smaller noise [8], [9]. The larger noises will yield more errors in a proportional way.…”

“…Lee and Kim [8] used reinforcement learning on 2D peg-in-hole simulation to develop an expert system. Gullapalli, Barto and Grupen [9] set up an associative reinforcement learning system based on the neural network. They let a 6-DOF manipulator learn to insert a peg by relationship between position and force sensing values, and output velocities.…”

Intelligent Control of Robotic Systems

“…A second group of learning methods, based on autonomous on-line learning procedures with the repetition of the working task, is also evaluated through several algorithms [105], [104], [152], [10], [88], [338]. The main distinction between these algorithms is in the aim of the learning, which is in the first case direct adjustment of control signals or parameters, while in the second case the aim of learning is the building of an internal model of the robotic system with compensation for system uncertainties.…”

Recent Advances in Robot Learning

“…These include the contact phase and lifting phases (Howe, et aI., 1990), regrasping, transport phases and release or insertion for assembly (Gullapalli, Barto & Grupen, 1994), each of which is indicated by specific sensory markers, such as tactile forces, that invoke differing control modes. An interesting problem is to learn reliable markers from this stream of sensory information that would indicate state transitions, as well as the correct control parameters that would hold during these states, as has been demonstrated is some initial work by (Pook & Ballard, 1993) using hidden Markov models.…”