Anticipatory grip force control using a cerebellar model

Abstract

Abstract-Grip force modulation has a rich history of research, but the results remain to be integrated as a neurocomputational model and applied in a robotic system. Adaptive grip force control as exhibited by humans would enable robots to handle objects with sufficient yet minimal force, thus minimizing the risk of crushing objects or inadvertently dropping them. We investigated the feasibility of grip force control by means of a biological neural approach to ascertain the possibilities for future application in robotics. As the cerebellum appears crucial for adequate grip force control, we tested a computational model of the olivo-cerebellar system. This model takes into account that the processing of sensory signals introduces a 100 ms delay, and because of this delay, the system needs to learn anticipatory rather than feedback control. For training, we considered three scenarios for feedback information: (1) grip force error estimation, (2) sensory input on deformation of the fingertips, and (3) as a control, noise. The system was trained on a data set consisting of force and acceleration recordings from human test subjects. Our results show that the cerebellar model is capable of learning and performing anticipatory grip force control closely resembling that of human test subjects despite the delay. The system performs best if the delayed feedback signal carries an error estimation, but it can also perform well when sensory data are used instead. Thus, these tests indicate that a cerebellar neural network can indeed serve well in anticipatory grip force control not only in a biological but also in an artificial system.

Specific Increases within Global Decreases: A Functional Magnetic Resonance Imaging Investigation of Five Days of Motor Sequence Learning

“…Sequence learning studies commonly describe extensive decreases in motor and cerebellar cortex, as well as frontal and parietal regions, as learning progresses Doyon, 2002, 2005;Floyer-Lea and Matthews, 2004;Lehéricy et al, 2005), and the model proposed by Doyon and Benali (2005) predicts decreases in most areas of the learning network. Decreases in M1 have been detected early in learning (Karni et al, 1995;Floyer-Lea and Matthews, 2004) and have been attributed to habituation or repetition suppression (Grill-Spector et al, 2006), whereas cerebellar cortical decreases have been interpreted as the decreased need for error correction as learning progresses (Ito, 2000;Penhune and Doyon, 2005;de Gruijl et al, 2009). The global decreases identified in our study are sharply contrasted by increases in specific subregions of M1 and cerebellar cortex.…”

Adaptive Robotic Control Driven by a Versatile Spiking Cerebellar Network

“…Learning was proposed to occur as long-term synaptic plasticity at the parallel fiber - Purkinje cell synapses in the form of long-term depression under instructive control by climbing fibers. Simulations carried out with these models suggested that a control scheme incorporating cerebellum adaptation played indeed a critical role for point-to-point arm movement and for prism glasses compensation in throwing at a target [11], [12], for anticipatory grip force modulation [13], and for Pavlovian collision avoidance [14]. While computational simulations guarantee repeatability over trials and therefore a systematic evaluation of the control scheme, testing with real robots is needed to assess the robustness and generalizability of models in closed-loop conditions, in which unwanted perturbations disturb learning and control.…”

Inferior Olive: All Ins and Outs

“…However, good results obtained in one specific task rarely carry over to another task without extensive tweaking of the model. For instance, a model that adequately learns the dynamics of a robot arm even without tight timing of error signals (Spoelstra et al 2000) may perform worse on a seemingly simpler task and require much tighter timing of error signals (De Gruijl et al 2009). Due to this inter-task variability and the necessity of task-specific tweaking of the model, the outcomes of large-scale models by themselves can be hard to interpret.…”