Safe Adaptive Switching among Dynamical Movement Primitives: Application to 3D Limit-Cycle Walkers

Sushant Veer, Ioannis Poulakakis

TL;DR

This paper introduces a control-theoretic framework that allows robots to safely switch between different movement patterns (like walking straight or turning) while facing persistent external disturbances. By calculating a mathematical "buffer time"—known as average dwell-time—the authors prove that a robot can adapt its behavior to follow a leader or compensate for pushes without losing stability or falling.

What Problem Does This Paper Address?

In robotics, complex behaviors are often built by stitching together simpler "Dynamical Movement Primitives" (DMPs)—think of these as pre-programmed skills like "walk straight" or "turn 30 degrees." While a robot might be perfectly stable while performing one skill, the act of switching between them can introduce instability, especially in unpredictable environments.

Most existing motion planners assume the environment is static or that disturbances are temporary. However, real-world robots face persistent forces, such as wind or a human collaborator pulling on them. If a robot switches behaviors too rapidly or at the wrong moment under these conditions, it risks leaving its "safe zone" (basin of attraction) and crashing. The challenge is to provide rigorous safety guarantees for switching that don't require the robot to know exactly what the external force is in advance.

Methods

The authors treat the robot as a switched system with multiple equilibria. Their approach relies on several key technical pillars:

  1. Lyapunov Stability & SoS Programming: They use Lyapunov functions to define "safe sets" for each movement primitive. To make this computationally practical for complex robots, they employ Sum-of-Squares (SoS) programming to estimate the Basins of Attraction (BoA)—the regions of the state space where the robot is guaranteed to remain stable.
  2. Average Dwell-Time (ADT): The core of the "safe switching" logic is a calculated lower bound on how long the robot must stay in one primitive before switching to the next. This ensures that the "energy" dissipated by the robot's stable dynamics outweighs any "energy gain" caused by the act of switching.
  3. Disturbance Agnosticism: A major contribution is Theorem 1, which proves that if the switching logic is safe for an unperturbed system, it remains safe under small, persistent disturbances. Crucially, the supervisor doesn't need to model the disturbance; it only needs to know the robot’s own stability properties.
  4. Hybrid Zero Dynamics (HZD): For the 3D bipedal walker example, the authors use HZD to project the robot's complex 17-dimensional dynamics onto a 2-dimensional "restricted" space. This massive reduction in dimensionality is what allows the sophisticated stability proofs to run efficiently.

Datasets

The paper does not use external datasets. Instead, it validates the framework using a high-fidelity simulation of a 3D limit-cycle bipedal robot. The model features nine degrees of freedom and is tested in a "collaborative transport" scenario where it must follow a leader's path based on interaction forces.

Key Insights and Ideas

  • The "Energy" Balance: Switching is viewed as an injection of uncertainty or "energy." The robot needs a "cooldown" period (the dwell-time) to settle back toward a stable state before another switch is allowed.
  • Safety through Inflation: The authors use a mathematical construction that "inflates" the safe sets of the primitives. If the union of these sets stays within the overall physical limits of the robot, safety is guaranteed.
  • Interaction Force as a Command: In the bipedal walking example, the robot doesn't know the leader's intended path. It treats the physical interaction force (the "push/pull") as a signal, integrating it over a stride to decide which walking primitive to switch to next.
  • Computational Efficiency: By combining HZD with SoS, the authors demonstrate that rigorous, formal safety proofs—often relegated to simple toy models—can be applied to complex, 3D walking robots.

Key Results

  • Safe Stride-to-Stride Switching: For the 3D walker, the calculated average dwell-time ($N_a$) was 0.99. This means the robot is stable enough to safely switch its walking primitive (e.g., from straight to turning) every single step.
  • Successful Collaboration: In simulations, the biped successfully followed a leader moving at 0.65 m/s along a curved path. Without the adaptive switching, the robot drifted away; with the framework, it remained within the "safe" grey region of its state space throughout the task.
  • Robustness: The framework proved effective even when the robot was subjected to persistent external forcing, maintaining stability without needing a precise model of those forces.

Limitations and Future Work

  • Small Disturbance Assumption: The mathematical proofs rely on the disturbances being "sufficiently small" (bounded by a value $\delta$). If a disturbance is massive (e.g., a heavy collision), the current safety guarantees may not hold.
  • Conservative Bounds: Dwell-time constraints can sometimes be conservative, potentially preventing the robot from switching as quickly as it physically could in an emergency.
  • Future Directions: The authors suggest extending this to more complex environments with obstacles and exploring larger disturbances where the "agnostic" assumption might need to be relaxed in favor of active disturbance estimation.

Summary updated

Abstract

Complex robot motions are frequently generated by composing simpler primitive movements. We use this approach to formulate robot motion plans as sequences of primitives to be executed one after the other. When dealing with dynamical movement primitives, besides accomplishing the high-level objective, planners must also reason about the effect of the plan's execution on the safety of the platform. This task is exacerbated by the presence of disturbances, such as non-vanishing external forces. To address this issue, we present a framework that builds on rigorous control-theoretic tools to generate safely executable motion plans for externally excited robotic systems. We illustrate the proposed framework on adapting the motion of a 3D bipedal robot model to persistent external forcing by switching among dynamic movement primitives, each corresponding to a limit-cycle walking gait.
BibTeX
@inproceedings{veer2019safe,
  author = {Sushant Veer and Ioannis Poulakakis},
  title = {{Safe Adaptive Switching among Dynamical Movement Primitives: Application to 3D Limit-Cycle Walkers}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2019},
  url = {https://openreview.net/forum?id=OhdXf050TZ},
}

Loading related content...

Figure