Multi-Agent Motion Planning on Industrial Magnetic Levitation Platforms: A Hybrid ADMM-HOCBF approach

1. Problem Statement

The paper addresses the challenge of real-time, safe, and scalable motion planning for holonomic multi-agent systems, specifically applied to the Beckhoff XPlanar industrial magnetic levitation platform.

• Context: The XPlanar system uses passive square movers actuated by magnetic tiles, allowing for high-precision, flexible 2D movement. Unlike traditional systems restricted to fixed tracks, these agents can move freely, creating a complex coordination problem.
• Core Challenge:
  • Scalability: Classical Centralized Model Predictive Control (MPC) solves a single optimization problem for all agents simultaneously. While optimal, its computational cost scales poorly (exponentially or high-order polynomially) with the number of agents ($N$), making it intractable for real-time control in large swarms.
  • Safety: Heuristic planners (e.g., A*, RRT) often lack formal safety guarantees or struggle with multi-agent deadlock. Conversely, standard distributed methods (like ADMM) often fail to guarantee safety in non-convex environments or for systems with higher relative degrees (e.g., acceleration inputs).
  • Goal: Develop a motion planner that is computationally efficient enough for real-time use with many agents while providing rigorous collision avoidance guarantees.

2. Methodology: Hybrid ADMM-HOCBF

The authors propose a hybrid framework that decouples the optimization of trajectories from the enforcement of safety constraints.

A. Decentralized Trajectory Optimization (ADMM)

To handle scalability, the authors use the Alternating Direction Method of Multipliers (ADMM) in a decentralized manner.

• Problem Reformulation: The global collision avoidance constraints are decoupled by introducing slack variables ($z_i$) and consensus constraints. Each agent solves a local optimization problem (OCP) considering only its own dynamics and actuation limits.
• Iterative Process:
  1. Agents solve local trajectory problems.
  2. Agents communicate positions to neighbors (in this implementation, all agents are neighbors to ensure bidirectional consistency).
  3. Agents update slack variables to enforce collision avoidance constraints locally.
  4. Dual variables are updated to drive consensus.
• Limitation: ADMM is only guaranteed to converge for convex problems. Since collision avoidance creates non-convex constraints, ADMM alone cannot formally guarantee safety or convergence to a collision-free solution.

B. Centralized Safety Filter (HOCBF)

To recover safety guarantees, the authors introduce a High-Order Control Barrier Function (HOCBF) layer.

• Mechanism: After the ADMM iterations produce a candidate control input, a secondary optimization step is performed.
• Formulation: A Quadratically Constrained Quadratic Program (QCQP) is solved centrally. This QCQP minimizes the deviation from the ADMM-generated input while enforcing the HOCBF condition:
  $$H_{i,j}(u_i, u_j) \geq 0$$
  This ensures that the relative distance between agents remains above a safety threshold, even for systems with acceleration inputs (relative degree $r=2$).
• Centralized Advantage: Unlike decentralized CBFs, the centralized safety filter on the XPlanar (which is controlled by a central computer) yields less conservative and more efficient trajectories.

C. Handling Complex Environments

• Rectangular Constraints: Agents are constrained within a rectangular arena using position limits.
• Corridor Strategy: For non-convex arenas (complex tile layouts), the environment is partitioned into overlapping rectangular corridors. Agents track a sequence of sub-targets (waypoints) to navigate through these corridors, allowing the use of simple rectangular constraints dynamically.

3. Key Contributions

1. Novel Hybrid Architecture: The combination of decentralized ADMM (for scalability and trajectory generation) with a centralized HOCBF (for rigorous safety guarantees) specifically tailored for systems with acceleration inputs.
2. Scalability vs. Safety Trade-off: Demonstrating that this hybrid approach achieves near-linear scalability ($O(N^{1.83})$) compared to the poor scalability of centralized MPC ($O(N^{2.96})$), while maintaining zero collisions.
3. Real-Time Industrial Validation: Development of a highly efficient C++ implementation (using FATROP and qpOASES solvers) and successful deployment on a physical Beckhoff XPlanar system with 5 moving agents.
4. Deadlock Resistance: The system demonstrates the ability to escape deadlocks through iterative ADMM refinement, even when initial reactive behaviors cause temporary stagnation.

4. Results and Performance

The paper presents extensive simulations and hardware experiments:

• Scalability Analysis:

  • Centralized MPC: Computation time scales as $O(N^{2.96})$. It becomes intractable for $N > 20$.
  • ADMM-HOCBF: Scales as $O(N^{1.83})$. The computation time is significantly lower for large $N$ (e.g., 30 agents).
  • Iteration Count: Increasing ADMM iterations ($m=20$) improves trajectory efficiency (lower transit time) but increases computation time. However, even with $m=20$, it outperforms centralized MPC for large swarms.

• Safety and Efficiency (5 Agents Simulation):

  • Collisions: 0 for all methods.
  • Transit Time: Centralized MPC was fastest (2.0s), followed by ADMM-HOCBF ($m=20$, 3.0s), and ADMM-HOCBF ($m=1$, 4.6s).
  • Computation Time: Centralized MPC took 1244ms (too slow for real-time). ADMM-HOCBF ($m=1$) took 115ms.
  • Safety Filter Activity: The HOCBF filter corrected inputs 65% of the time for $m=1$ and 36% for $m=20$, proving the filter is active and necessary.

• Hardware Experiments (XPlanar):

  • Real-Time Performance: The C++ implementation achieved an average computation time of 20 ms for $m=1$ and 204 ms for $m=20$ with 5 agents, confirming real-time capability.
  • Dynamic Scenarios: Successfully handled continuous operation with changing targets and navigation through non-convex arenas using the corridor strategy.

5. Significance

This work bridges the gap between theoretical multi-agent control and industrial application.

• Industrial Relevance: It proves that complex, flexible, multi-agent logistics (like warehouse automation) can be managed in real-time without sacrificing safety.
• Methodological Advance: It resolves the classic conflict between the computational efficiency of distributed methods and the safety guarantees of centralized control. By offloading safety to a centralized filter, the system retains the scalability of distributed planning.
• Future Impact: The approach paves the way for larger swarms of autonomous agents in dynamic manufacturing environments, moving beyond fixed tracks to fully flexible, collision-free navigation.