Learning Task Constraints from Demonstration for Hybrid Force/Position Control

Abstract

We present a novel method for learning hybrid force/position control from demonstration. We learn a dynamic constraint frame aligned to the direction of desired force using Cartesian Dynamic Movement Primitives. In contrast to approaches that utilize a fixed constraint frame, our approach easily accommodates tasks with rapidly changing task constraints over time. We activate only one degree of freedom for force control at any given time, ensuring motion is always possible orthogonal to the direction of desired force. Since we utilize demonstrated forces to learn the constraint frame, we are able to compensate for forces not detected by methods that learn only from the demonstrated kinematic motion, such as frictional forces between the end-effector and the contact surface. We additionally propose novel extensions to the Dynamic Movement Primitive (DMP) framework that encourage robust transition from free-space motion to in-contact motion in spite of environment uncertainty. We incorporate force feedback and a dynamically shifting goal to reduce forces applied to the environment and retain stable contact while enabling force control. Our methods exhibit low impact forces on contact and low steady-state tracking error.

Quick Setup of Force-Controlled Industrial Gluing Tasks Using Learning From Demonstration

“…Constraint frames are usually chosen manually based on the task specifications Raibert and Craig (1981). Common choices include the tool frame (Kronander and Billard, 2014;Peternel et al, 2017) and the surface normals (Deng et al, 2016;Conkey and Hermans, 2019). The latter can also be estimated, e.g.…”

“…from the velocity of the contact point (Karayiannidis et al, 2014;Sloth and Iturrate, 2021). Conkey and Hermans (2019) present a framework for learning a dynamic constraint frame aligned to the direction of desired force using CDMPs and force control to ensure that the force is aligned in the normal plane with the direction of the movement. In (Kober et al, 2015), multiple demonstrations of a task are used to segment a series of separate movement primitives, each associated with a specific discrete frame of reference.…”

“…Our work is conceptually similar to Kronander and Billard (2012); Kober et al (2015); Lee et al (2015); Conkey and Hermans (2019), in that we are interested in learning task constraints from demonstration, such that we can then adapt the parameters/gains of a compliant controller to the structure of the task. However, in contrast to other approaches (Kronander and Billard, 2012;Kober et al, 2015;Lee et al, 2015), which require either multiple (typically more than three) demonstrations of the task or multiple execution iterations in order to determine these constraints, our approach can do so from a single demonstration by exploiting the surface normal estimator to determine the task geometry.…”

“…Our work is methodologically most similar to Conkey and Hermans (2019). While their method is also able to learn task constraints from a single demonstration, they rely on the forces observed during the demonstration to do so.…”

“…As indicated above, in order to reduce contact force transients during the execution, we modulate the goal of the DMP slowly into the contact surface. Although this method is similar to Conkey and Hermans (2019), our use of it is qualitatively different. Whereas Conkey and Hermans (2019) use goal modulation reactively to adapt to unforseen generalizations of the task-i.e., a contact that is established either earlier or later than anticipated-we do so proactively by purposely setting the initial goal higher than originally demonstrated, to ensure that impact forces are always reduced.…”