Impact-Aware Robotic Manipulation: Quantifying the Sim-To-Real Gap for Velocity Jumps

TL;DR

This paper introduces a method to generate and validate impact maps for robotic manipulation using physics simulations and experiments, enabling better modeling of velocity jumps during impacts for improved control and planning.

Contribution

It proposes an experimental validation approach for impact maps derived from simulations, incorporating a reference spreading control framework to account for complex impact scenarios.

Findings

Achieved a 3.1% average error between simulated and experimental post-impact velocities.

Validated impact maps across complex impact scenarios using the proposed framework.

Demonstrated the effectiveness of the impact map validation in real robotic experiments.

Abstract

Impact-aware robotic manipulation benefits from an accurate map from ante-impact to post-impact velocity signals to support, e.g., motion planning and control. This work proposes an approach to generate and experimentally validate such impact maps from simulations with a physics engine, allowing to model impact scenarios of arbitrarily large complexity. This impact map captures the velocity jump assuming an instantaneous contact transition between rigid objects, neglecting the nearly instantaneous contact transition and impact-induced vibrations. Feedback control, which is required for complex impact scenarios, will affect velocity signals when these vibrations are still active, making an evaluation solely based on velocity signals as in previous works unreliable. Instead, the proposed validation approach uses the reference spreading control framework, which aims to reduce peaks and jumps in the control feedback signals by using a reference consistent with the rigid impact map together with a suitable control scheme. Based on the key idea that selecting the correct rigid impact map in this reference spreading framework will minimize the net feedback signal, the rigid impact map is experimentally determined and compared with the impact map obtained from simulation, resulting in a 3.1% average error between the post-impact velocity identified from simulations and from experiments.