Learning Interactive Behaviors for Musculoskeletal Robots Using Bayesian Interaction Primitives

Abstract

Musculoskeletal robots that are based on pneumatic actuation have a variety of properties, such as compliance and back-drivability, that render them particularly appealing for human-robot collaboration. However, programming interactive and responsive behaviors for such systems is extremely challenging due to the nonlinearity and uncertainty inherent to their control. In this paper, we propose an approach for learning Bayesian Interaction Primitives for musculoskeletal robots given a limited set of example demonstrations. We show that this approach is capable of real-time state estimation and response generation for interaction with a robot for which no analytical model exists. Human-robot interaction experiments on a 'handshake' task show that the approach generalizes to new positions, interaction partners, and movement velocities.

Human-Robot Handshaking: A Review

“…More recent works model reaching using machine learning. Campbell et al [14] use imitation learning to learn a joint distribution over the actions of the human and the robot. During testing time, the posterior distribution is inferred from the human's initial motion from which the robot's trajectory is sampled.…”

“…Major Findings Modelling of reaching behaviours draws heavily on learning from human interactions, unlike other robotic grasping/manipulation tasks, where a lot of it can be learnt from scratch. This provides a strong prior to help make the motions more human-like and can also be used to initialise [14] or guide [17] the learning.…”

Learning Human-like Hand Reaching for Human-Robot Handshaking

“…Vinayavekhin et al [9] aim to bridge this gap in the context of human-robot handshaking using a recurrent network for predicting the human-hand motion and devise a simple controller for the robot's response motion. Such interaction dynamics during human-robot handshaking are also captured implicitly by Campbell et al [10] who learn a joint distribution over the trajectories of the human and the robot. However, their approach is robot-specific and would need to be re-trained with human interaction partners when applied to new robots.…”

Learning Whole-Body Human-Robot Haptic Interaction in Social Contexts

“…Our estimate of the underlying latent model contains correlated uncertainties between the individual weights, due to a shared error in the phase estimate. Intuitively, this is because a temporal error induces a correlated error in spatial terms due to the phase dependency of the basis functions [4]. Probabilistically, we represent this with the augmented state vector s = [φ, φ, w] and the following definition:…”

“…We model the interaction as a Bayesian Interaction Primitive (BIP) [2], [3], [4], a spatiotemporal LfD framework. This model is capable of predicting both an appropriate robotic response (consisting of joint trajectories) as well as contact forces that should be exerted by the robot, given observations of the partner's pose and the force currently exerted by the partner onto the robot.…”