Accepted Works

Authors:

Zixi Liu and Robert D. Howe

Abstract:

Reliable grasping and manipulation for deformable objects require accurate contact modeling and grasp stability estimation. One key component in contact modeling and stability estimation is the coefficient of friction, which is typically estimated as a standard single value from the literature, even though its actual values are variable and depend on many factors, such as compliance and contact velocity, especially for deformable objects. These factors are often overlooked in the applications of analytical grasp models as well as machine learning methods. Here we compare the coefficient of friction of objects with different compliance but identical materials (thicker/multi-layered vs. thinner/single-layered) at varying contact velocity, using a highly instrumented robot hand and vision-based tracking on both the robot hand and the object. The results show that compliance, as well as contact velocity, affect the coefficient of friction and stability estimation using grasp analysis. These findings suggest that reliable grasping and manipulation, whether from analytical grasp models or machine learning methods, require the ability to sense friction. This also implies that machine learning methods will require inputs from friction sensing. Without friction sensing capabilities, robotic grasping and manipulation are constrained to a much narrower range of objects.

Authors:

Federico Ulloa, Wilm Decré, Erwin Aertbeliën, and Herman Bruyninckx

Abstract:

This paper presents a methodology for robust insertion skill definitions, that allows developers to exploit their insights in (i) the (active and passive) compliance properties of both robot and environment, and (ii) the source and magnitude of geometrical uncertainties. The core ideas behind the methodology are (i) it is always easier to deal with geometrical uncertainties one degree of freedom at a time than several together, and (ii) the passive compliance in the assembly parts can be used as a feature instead of as a bug, most notably in the case of engineered compliances such as snap fits.

Authors:

Junyoung Kim, Wonbum Yun, and Sehoon Oh

Abstract:

For the environment-interactive compliant robots, the series elastic actuator (SEA) is concerned as a promising actuation method. With the serially connected spring, SEA can measure interaction torque as a sensor and absorb shocks introduced from the interactions to protect the transmission of the actuator. To develop SEA, designing a sufficiently soft spring may be required to achieve precise torque measurement. However, since soft springs generally cannot hold higher torque, it is difficult to use softer springs for SEA, and designing the spring to meet the requirements of the actuator often needs a complex process. In this study, we proposed the design of a novel spring structure which can provide high torque capacity with a wide range of stiffness algebraic analyses. Also, we proposed the structure and mechanical design of SEA to consider the characteristics of the proposed spring. The specification and quality of the proposed spring and actuator are evaluated through a series of experiments.

Authors:

Julius Jankowski and Sylvain Calinon

Abstract:

Setting the right robot compliance for a given task can be difficult. Typically, a robot should have high physical compliance such that it does not generate high forces when in unforeseen contact with its environment. Moreover, it lets it react smoothly to unforeseen changes in the motion plan. In contrast, e.g., tight tolerance tasks demand for high precision in the robot execution, which - due to model imperfections - requires high positional feedback gains, i.e. high physical stiffness. Probabilistic Adaptive Control addresses this trade-off via time-varying feedback gains. They are inferred in an online-manner by fusing information about a) asymptotically stable tracking of deterministic trajectories, b) closed-loop tracking uncertainty, e.g. due to modeling errors, and c) task-specific desired precision instructed through kinematic demonstrations.

Authors:

Thomas Baaij*, Marn Klein Holkenborg*, Maximilian Stölzle*, Daan van der Tuin*, Jonatan Naaktgeboren, Robert Babuska, and Cosimo Della Santina

Abstract:

While soft fingers are an interesting approach for compliant manipulation, sensing their shape without obstructing their movements and modifying their natural softness requires innovative solutions. This abstract proposes to use magnetic sensors fully integrated into soft continuum structures to achieve proprioception. Magnetic sensors are compact, sensitive, and easy to integrate into a soft robot. We then propose a neural architecture to make sense of the highly nonlinear relationship between the perceived intensity of the magnetic field and the shape of the robot. By injecting a priori knowledge from the kinematic model, we obtain a precise learning strategy yet not data intensive. We verify our approach in experiments involving one soft segment containing a cylindrical magnet and three magnetoresistive sensors. We achieve mean relative errors of 17.4%, and as low as 7% by training on more diverse datasets.

Authors:

Ka Hei Mak and Jungwon Seo

Introduction:

The presented study addresses how to perform the manipulation of scooping over a wide range of objects at high speed, termed as high-speed scooping. This is a departure from existing solutions to robotic scooping through underactuation, with a limited ability to actively adapt to various situations, or motion control, with dependence on accurate geometric information. Critical to high-speed scooping are the capabilities for rapidly responding to contact interactions between the object, robot, and environment under errors and uncertainties regarding the geometric and physical properties of the bodies that physically interact.