Safety-Critical Coordination for Cooperative Legged Locomotion via Control Barrier Functions

Jeeseop Kim, Jaemin Lee, Aaron D. AmesIROS 2023 · DOI

TL;DR

This paper introduces a novel safety-critical framework using Control Barrier Functions (CBFs) to enable cooperative legged robots to navigate complex environments while maintaining formations and avoiding obstacles. The approach uses a hierarchical controller that guarantees safe trajectories and coordinated motion, demonstrated through simulations and real-world experiments with quadrupedal robots.

What Problem Does This Paper Address?

Deploying teams of legged robots for tasks like collaborative object transportation in human-centered environments presents significant challenges. These robots need to move dynamically, maintain specific formations (holonomic constraints), and simultaneously avoid collisions with each other and their surroundings. Synthesizing control algorithms that guarantee safety under these complex, high-dimensional, and hybrid dynamics is particularly difficult. Existing methods often struggle to ensure safety in the presence of these interconnected constraints and dynamic behaviors, especially when dealing with multiple robots. This paper aims to bridge this gap by providing a robust safety-critical coordination strategy for such systems.

Methods

The paper proposes a three-layered control architecture:

  1. Safety-Critical Planner (Top Layer): This layer uses Control Barrier Functions (CBFs) to generate safe, reduced-order kinematic trajectories for the cooperative system. CBFs are mathematical tools that guarantee a system remains within a "safe set" by ensuring a specific function (representing safety) remains non-negative. This layer extends CBF theory to handle multiple agents and holonomic constraints, ensuring collision avoidance and formation maintenance.
  2. Trajectory Planner with Interconnected SRB Dynamics (Middle Layer): This layer uses Model Predictive Control (MPC) with interconnected Single Rigid Body (SRB) dynamics. It takes the safe trajectories from the top layer and computes optimal Ground Reaction Forces (GRFs) to track these trajectories, while accounting for the dynamic interactions between robots due to holonomic constraints.
  3. QP-based Virtual Constraint Controller (Bottom Layer): This layer uses a distributed Quadratic Programming (QP) controller to generate whole-body motion for each individual robot. It utilizes the full-order nonlinear model of each agent and ensures that the robots follow the prescribed GRFs and velocities from the upper layers, while also satisfying virtual constraints for coordinated motion.

A key innovation is the extension of CBFs to handle both obstacle avoidance and the holonomic constraints between robots, formulated as a safety-critical coordination problem solvable via QP. The paper also incorporates Input-to-State Safety (ISSf) CBFs to ensure robustness against bounded control input perturbations.

Datasets

Not discussed in the paper. The evaluation relies on numerical simulations and hardware experiments using quadrupedal robots.

Key Insights and Ideas

  • Unified Safety Framework: The paper successfully unifies obstacle avoidance and holonomic constraint satisfaction within a single CBF-based safety framework. This is crucial for ensuring robust cooperative behavior.
  • Hierarchical Control for Complexity: The three-layered architecture effectively breaks down the complex problem into manageable sub-problems, allowing for the integration of different modeling approaches (kinematic, SRB dynamics, full-order nonlinear) and control techniques (CBF, MPC, QP).
  • ISSf for Robustness: The incorporation of Input-to-State Safety (ISSf) CBFs demonstrates the framework's resilience to control input noise, a critical factor for real-world deployment.
  • Experimental Validation: The successful demonstration on physical quadrupedal robots, not just simulations, significantly validates the practical applicability of the proposed method.

Key Results

  • The proposed safety-critical planner successfully enabled cooperative locomotion for two quadrupedal robots while avoiding two, three, and four obstacles in simulation.
  • The distance between agents remained close to the imposed constraint, even with input noise, demonstrating the effectiveness of the ISSf CBFs.
  • Both simulation and hardware experiments showed that the holonomically constrained full-order systems could successfully track safety-ensured trajectories without losing stability.
  • The solving time for the QP-formulated safety-critical planner was reported to be approximately 0.9-1.3 ms for two to four obstacles, indicating real-time feasibility.

Limitations and Future Work

  • Extension to More Agents: The authors suggest exploring how the framework can be extended to a greater number of agents.
  • 3D Locomotion: Future work includes extending the safety-critical planner to three dimensions to handle varying terrain types.
  • Uncertain Environments: The paper acknowledges that future work could explore more complex and uncertain environments.

Summary updated

BibTeX
@inproceedings{kim2023safety,
  author = {Jeeseop Kim and Jaemin Lee and Aaron D. Ames},
  title = {{Safety-Critical Coordination for Cooperative Legged Locomotion via Control Barrier Functions}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  doi = {10.1109/iros55552.2023.10341987},
}
Figure