Verifying Safe Transitions between Dynamic Motion Primitives on Legged Robots

Wyatt Ubellacker, Noel Csomay-Shanklin, Tamás G. Molnár, Aaron D. Ames

TL;DR

This paper introduces a formal framework for ensuring safe transitions between dynamic motion primitives on legged robots. By constructing a "motion primitive graph," the system can plan and execute complex sequences of behaviors, like walking and jumping, without violating safety constraints, a crucial step towards more robust and functional autonomous robots.

What Problem Does This Paper Address?

Modern autonomous systems, especially robots, often rely on combining discrete behaviors (like "walk," "stand," "jump") to achieve complex tasks. While this approach works well for simpler, quasi-static robots, it becomes challenging for dynamic systems like legged robots. The core problem is ensuring that when a robot switches from one dynamic behavior to another, it does so safely. This means not only reaching the target behavior but also avoiding unsafe states during the transition itself, which could lead to falls or damage. Existing methods often treat these transitions as instantaneous or don't explicitly account for the robot's dynamic state, leading to potential failures. This paper aims to formalize and verify these transitions for dynamic legged robots.

Methods

The paper's main contribution is a framework for building a "motion primitive graph." Here's a breakdown of their approach:

  • Motion Primitives: They define a motion primitive as a dynamic behavior characterized by a setpoint (desired trajectory), a control law, a safe region of attraction (RoA), and an explicit safe RoA. They categorize primitives into fixed, periodic, and transient.
  • Safety Oracle: To verify if a transition from primitive A to primitive B is safe, they introduce a "safety oracle." This oracle checks if the robot's state, starting from the setpoint of primitive A at a given time, remains within the safe set of primitive B for a specified horizon and eventually converges to primitive B's explicit safe RoA.
  • Motion Primitive Graph Construction: They use the safety oracle to systematically check all possible pairs of motion primitives. If a safe transition is possible, an edge is added to the graph. Class 1 transitions (always possible) are represented by solid edges, and Class 2 transitions (possible only at specific times) by dashed edges. This graph then allows for standard graph search algorithms to find safe transition paths.
  • Hardware Implementation: The framework was applied to a Unitree A1 quadruped robot, implementing primitives like Lie, Stand, Walk (various speeds), and Jump.

Datasets

Not discussed in the paper. The authors generated their own motion primitives and tested them on a physical robot.

Key Insights and Ideas

  • Formalizing Dynamic Transitions: The paper provides a rigorous mathematical framework for defining and verifying safe transitions between dynamic behaviors, moving beyond heuristic approaches.
  • The "Safety Oracle" Concept: This is a clever way to abstract the complex verification problem. Instead of needing a full analytical solution for the region of attraction, they use a time-bound simulation check, making it computationally tractable.
  • Motion Primitive Graph as a Planning Tool: The resulting graph acts as a high-level "state machine" that guarantees safety. This allows for offline planning of complex maneuvers, which can then be executed online.
  • Class 1 vs. Class 2 Transitions: Distinguishing between always-safe and time-dependent safe transitions is crucial for real-world applications, as it reflects the inherent dynamics of the system.
  • The Importance of Intermediate Primitives: The experiments highlight that direct transitions are often unsafe. The graph-based approach naturally incorporates necessary intermediate primitives (like "Stand" or "Walk(in place)") to bridge unsafe gaps.

Key Results

The paper demonstrates the successful application of their framework on a Unitree A1 quadruped.

  • Safe Execution of Complex Sequences: They successfully executed a sequence of desired primitives (Lie, Walk(fast), Jump, Land, Stand) that would have failed with a "naïve" transition strategy.
  • Violation of Safety Constraints Avoided: The naïve approach led to violations of the no-slip safety constraint and failure to track desired behaviors. The proposed method successfully avoided these issues.
  • Experimental Validation: The results were validated through hardware experiments, showing the practical applicability of the framework.

Limitations and Future Work

  • Conservativeness: The safety oracle's verification is sufficient but not necessary. This means a "False" from the oracle doesn't definitively mean a transition is impossible, just that their specific check failed. They suggest that increasing the horizon T can reduce this conservativeness.
  • Real-world Uncertainty: The current framework assumes a perfect dynamic model. Future work aims to address real-world uncertainties and disturbances.
  • Weighted Graphs: For systems with many primitives, the graph can become large. They propose exploring weighted motion primitive graphs to optimize for metrics like transition duration or peak torque.
  • Arbitrary Initial States: The current method focuses on transitions from a primitive's setpoint. Extending this to handle states outside the safe region of attraction is a future research direction.

Summary updated

Abstract

Functional autonomous systems often realize complex tasks by utilizing state machines comprised of discrete primitive behaviors and transitions between these behaviors. This architecture has been widely studied in the context of quasi-static and dynamics-independent systems. However, applications of this concept to dynamical systems are relatively sparse, despite extensive research on individual dynamic primitive behaviors, which we refer to as "motion primitives." This paper formalizes a process to determine dynamic-state aware conditions for transitions between motion primitives in the context of safety. The result is framed as a "motion primitive graph" that can be traversed by standard graph search and planning algorithms to realize functional autonomy. To demonstrate this framework, dynamic motion primitives -- including standing up, walking, and jumping -- and the transitions between these behaviors are experimentally realized on a quadrupedal robot.
BibTeX
@inproceedings{ubellacker2021verifying,
  author = {Wyatt Ubellacker and Noel Csomay-Shanklin and Tamás G. Molnár and Aaron D. Ames},
  title = {{Verifying Safe Transitions between Dynamic Motion Primitives on Legged Robots}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2021},
  doi = {10.1109/iros51168.2021.9636537},
}
Figure