Constrained Dynamics Simulation: More With Less

TL;DR

Efficient constrained dynamics simulation is central to robot control, learning-based methods, and planning, but current solvers are bottlenecked by high-cost inner dynamics. The work revisits and extends linear-complexity CDAs, reviving PV-CDA with $O(n + m^2 d + m^3)$ complexity and extending to floating-base trees, while PV-soft and PV-early realize asymptotic $O(n+m)$. It further introduces PV-OSIM, PV-OSIMr, and cABA-OSIM with complexities $O(n+m^2d+m^2)$, $O(n+m^2)$, and $O(n+m^2)$, respectively, achieving up to $3\times$ speedups over Featherstone's LTL-OSIM and implemented in Pinocchio for practical use. The paper also outlines a path toward differentiable, data-efficient simulation to enable whole-body MPC and RL on resource-constrained hardware, with a long-term goal of democratizing robotics research.

Abstract

Efficient robot dynamics simulation is a fundamental problem key for robot control, identification, design and analysis. This research statement explores my current progress in this field and future research directions.