Impact-Aware Task-Space Quadratic-Programming Control

Abstract

Generating on-purpose impacts with rigid robots is challenging as they may lead to severe hardware failures due to abrupt changes in the velocities and torques. Without dedicated hardware and controllers, robots typically operate at a near-zero velocity in the vicinity of contacts. We assume knowing how much of impact the hardware can absorb and focus solely on the controller aspects. Hybrid controllers with reset maps provided elegant solutions for given impact tasks. The novelty of our approach is twofold: (i) it uses the task-space inverse dynamics formalism that we extend by seamlessly integrating impact tasks; (ii) it does not require separate models with switches or a reset map to operate the robot undergoing impact tasks. Our main idea lies in integrating post-impact states prediction and impactaware inequality constraints as part of our existing general-purpose whole-body controller. To achieve such prediction, we formulate task-space impacts and its spreading along the kinematic and potentially arborescent, structure of a floating-base robot with subsequent joint velocity and torque jumps. As a result, the feasible solution set accounts for various constraints due to expected impacts. In a multi-contact situation of under-actuated legged robots subject to multiple impacts, we also enforce the dynamic equilibrium margins. By design, our controller does not require precise knowledge of impact location and timing. We assessed our formalism with the humanoid robot HRP-4, generating maximum contact velocities, neither breaking established contacts nor damaging the hardware.

Humanoid Loco-Manipulations Using Combined Fast Dense 3D Tracking and SLAM With Wide-Angle Depth-Images

“…Indeed, in order to proceed faster the robot shall be able to make contact with non-zero relative velocities, which we are prohibiting for the time being. Our recent results [58] allow extensions to our whole body control to handle impacts, yet it still has to be integrated to the mc_rtc framework. • Lack of whole body preview: our task-space QP controller is local and hence agnostic to task-induced dynamics in future iterations (except for walking).…”

Learning to Pick at Non-Zero-Velocity from Interactive Demonstrations

“…As a final remark, it is worth underlining that the capability of correcting the orientation after the demonstration was not enabled due to the limitations of the teleoperation interface, not due to any limitations surrounding the algorithm itself and is thus left to future work. 1 https://youtu.be/6brynHStxGE…”

“…We as humans, on the other hand, tend to grasp things in a single fluent and quick motion. Of course, robots should also be able to complete a task fairly quickly, which in the case of grasping introduces a number of challenges, both from a control point of view [1] as well as a learning point of view [2].…”

Human inspired fall arrest strategy for humanoid robots based on stiffness ellipsoid optimisation

“…To minimise damage, falling trajectory optimisation [15,16], pose reshaping [17] and adaptive compliance control [18,19] have been proposed when a fall is inevitable. To generate on-purpose impacts with rigid objects, Wang et al proposed an impact-aware [20] control method with task-space quadratic optimisation. These mentioned strategies mainly focus on how to reduce damage during contact.…”