Title:
Anticipatory and Reactive Motion Planning
Anticipatory and Reactive Motion Planning
Author(s)
Scheele, Samuel
Advisor(s)
Ravichandar, Harish
Chernova, Sonia
Gombolay, Matthew
Chernova, Sonia
Gombolay, Matthew
Editor(s)
Collections
Supplementary to
Permanent Link
Abstract
Close-proximity human-robot interaction requires that robots plan motions that both efficiently perform tasks and adapt to human behavior and preferences. However, these requirements are mediated by a number of potentially conflicting aspects of human-robot interactions. This thesis proposes an anticipatory framework for simultaneously optimizing robot trajectories for multiple cost functions, then makes refinements to the framework using techniques from nonlinear model-predictive control to make it both anticipatory and reactive. We leverage recent advances in human motion prediction to plan adaptive robot motion in anticipation of the human partner’s movements.
Our approaches explicitly model and allow users to balance cost functions related to various aspects of an interaction. We illustrate the effectiveness of our framework by simultaneously optimizing for distance between the robot and user, the legibility of the robot’s trajectory, the tendency of the robot to stay in the user’s field of view, the robot’s speed, and deviation from the default trajectory, based on anticipated and current human motion. These cost functions are intended to balance safety and user comfort against task efficiency in a Close-Proximity Human-Robot Collaboration (CP-HRC) setting.
We introduce CoMOTO, our method for simultaneously optimizing these objectives offline, and show that it typically generates solutions at least as good as baseline methods, but works across a wider variety of interaction profiles. While CoMOTO takes anywhere from several seconds to several minutes to optimize a trajectory, we also introduce a Nonlinear Model-Predictive Control (NMPC) approach, RT-CoMOTO, which is capable of running in real time due to our use of receding-horizon optimization, more recent techniques in solving constrained nonlinear optimization problems, and reformulations of cost functions. Our results in simulation demonstrate that RT-CoMOTO performs comparably to offline planning in terms of comfort-related metrics, while running about 400x faster than CoMOTO in our test cases. Finally, we implement RT-CoMOTO on a Jaco Gen3 manipulator and test it against a number of baseline HRI methods in a user study. Preliminary results suggest that RT-CoMOTO plans efficient motion and is viewed favorably by human partners relative to existing methods in robot motion planning and HRI.
Sponsor
Date Issued
2023-05-02
Extent
Resource Type
Text
Resource Subtype
Thesis