Multi-contact motion retargeting from human to humanoid robot

Abstract

We propose a framework for real-time online and offline retargeting of a human actor motion to a humanoid robot motion involving multi-contact configuration changes between the human/humanoid and their environments. The framework is based on the specification within a multi-contact QP control formulation of tracking tasks for a selected set of body-segments/links, the ones either used for a manipulation task from a fixed multi-contact stance or susceptible to be used as contact supports or as locomotion supports on the environment in the course of the execution of the motion. The framework is applied in an online setting for simultaneous human-robot motion tracking (in the case of no contact configuration change) and in an offline setting for tracking the playback of the recorded human motion that is pre-processed to extract from it the sequence of contact change events as a necessary motion information exploited by the tracking algorithm. We present a real robot experiment with HRP-4 for the online setting and a dynamics simulation experiment for the offline setting to validate the proposed approach

Adaptive whole-body manipulation in human-to-humanoid multi-contact motion retargeting

“…TRACKING QP In [7], human-to-humanoid multi-contact motion retargeting with contact changes is addressed in a two-stage approach. In the first stage, the sequence of contact change events…”

“…The constraints for torque limits, joint angle/velocity limits, and collision avoidance (6) are described in [7], and are excluded here to emphasize the other equations which are dependent on the tracking/contact sets partition (E 1 , E 2 ).…”

“…Similarly, if the robot is about to release a manipulated object, it must first check that its individual (not multi-robot) CoM lies within its support polygon. For contact additions, we add two temporary end effector tasks to ensure a stable contact, as described in [7]: a surface vector stabilization task to ensure that the robot's end-effector surface is parallel to the object surface, and a surface transform task to drive the distance between the two surfaces to zero. Once the contact change conditions are fulfilled, the robot resumes tracking the demonstration.…”

“…To deal with multiple (possibly non-coplanar) contacts on arbitrary links, our work builds on that of Di Fava et al [7], which uses a QP controller to track human motion with Cartesian space "tasks" and to reason explicitly about contact friction cones [8]. Early QP control work in computer graphics demonstrated multi-objective control for generating natural movements in animated characters [9] [10].…”

“…In Section II we summarize multi-contact motion retargeting with a QP controller, as introduced in [7]. In Section III we introduce additionally the new multi-robot QP formulation for multicontact motion retargeting.…”