Researchers have developed a hybrid control framework that combines predictive planning using ADMM (Alternating Direction Method of Multipliers) with safety-certified Control Barrier Functions (CBFs) to enable self-assembly and reconfiguration of modular robotic boats. This approach resolves collisions and deadlocks in real-time, running at 5 Hz on onboard microcomputers.
What the Researchers Built
The team constructed a distributed motion planning system for fleets of self-reconfiguring robotic boats—modular vessels that can autonomously form arbitrary shapes (e.g., rectangles, triangles) and then break apart and reconfigure into new formations. Each boat has a nonholonomic unicycle kinematic model with velocity constraints, and the boats communicate only with their immediate neighbors in a communication graph.
The framework has three layers: a centralized task-level planner that assigns target positions for the desired shape, a distributed predictive control layer using ADMM that generates collision-free trajectories over a finite horizon, and a reactive safety filter based on Control Barrier Functions that provides instantaneous collision avoidance guarantees. The CBF filter runs at a higher frequency than the ADMM planner, overriding trajectories that would lead to collisions. This hybrid design was validated both in simulation and on physical hardware using small-scale robotic boats.
Key Results
In ablation studies, the researchers compared their hybrid ADMM-CBF framework against two baseline approaches: ADMM-only (predictive planning without reactive safety filter) and CBF-only (purely reactive obstacle avoidance without predictive planning). The ADMM-only approach suffered collisions even with small swarm sizes because the multi-agent collision avoidance problem is highly nonconvex, and distributed ADMM limited to a single iteration per control step often fails to converge to a safe consensus. The CBF-only approach avoided collisions but frequently caused deadlocks where boats would get stuck oscillating near goal positions, unable to navigate around each other.
The hybrid framework successfully achieved collision-free, deadlock-free formation control for swarms up to the tested size, running at 5 Hz on onboard microcomputers. Physical experiments with small-scale boats confirmed the simulation results, demonstrating smooth reconfiguration in a water tank. The computational cost scales linearly with the number of agents, making the approach suitable for real-time operation.
How It Works
The system operates on a hierarchical control architecture. At the top level, a centralized algorithm assigns each boat a target position in the desired formation. These targets are then sent to the distributed control layer.
The distributed layer uses ADMM to solve a decentralized Model Predictive Control (MPC) problem. Each boat computes its own planned trajectory over a prediction horizon, while sharing its plan with neighbors. The ADMM algorithm iterates to find a consensus among modules—seeking trajectories that avoid collisions while converging to goal positions. However, because the algorithm is limited to a single iteration per control step to meet real-time constraints, it cannot guarantee safety alone.
This is where the CBF safety filter comes in. A Control Barrier Function is a mathematical certificate that ensures a system stays within a safe set (e.g., minimum distance between boats). The CBF filter is implemented as a Quadratic Program (QP) that minimally modifies the velocity command from the ADMM planner—if that command would lead to a violation of the safety barrier, the filter overrides it with the nearest safe velocity. The CBF runs at a higher update rate (e.g., 50 Hz) than the ADMM planner (5 Hz), providing continuous safety between planner updates.
Additionally, the framework includes a distance adaptation mechanism: when two modules are assigned to close positions, the CBF constraints dynamically shrink the safe distance threshold to allow the formation to tighten, preventing deadlocks that would occur with a fixed safety margin.
Why This Matters for Robotics
Self-reconfiguring robotic systems have long been a holy grail for applications like adaptive infrastructure, environmental monitoring, and collective transport. This work brings that vision closer to practicality by providing safety guarantees and deadlock resolution with distributed computation—critical for scaling to swarms of dozens or hundreds of autonomous units.
The hybrid ADMM-CBF architecture is particularly valuable because it separates the problem into two tractable parts: predictive planning (which can be done at a lower frequency with incomplete convergence) and reactive safety (which runs fast and provides guarantees). This same architectural pattern could be applied to other multi-robot coordination problems, such as warehouse robots or used industrial robots performing collaborative tasks in shared workspaces. The approach demonstrates that you can have both real-time performance and formal safety while maintaining distributed computation.
Limitations and Open Questions
The current framework has two main limitations. First, it assumes a centralized assignment of target positions, which becomes a bottleneck for much larger swarms. Second, the CBF-based safety filter requires an accurate model of each boat's dynamics—in practice, water currents and wave disturbances can violate that model, causing the filter to compute unsafe override commands. The researchers tested only in calm water conditions.
Additionally, scalability testing revealed that beyond a certain swarm size (not reported in the available text), the ADMM convergence rate degrades, leading to more frequent reliance on the CBF filter and potential deadlocks. Finally, all experiments used rectangular shapes; more complex target formations with interior holes or non-convex shapes may introduce new coordination challenges. Future work aims to address collective transport, navigation in constrained waterways, and robustness to environmental disturbances.
Frequently Asked Questions
What is a Control Barrier Function (CBF)? A CBF is a mathematical function that defines a safety region (e.g., minimum distance between boats). A controller can be modified using a quadratic program to ensure the system never leaves that safe region, providing real-time collision avoidance guarantees.
How does ADMM differ from other distributed optimization methods? ADMM decomposes a large optimization problem into smaller subproblems solved in parallel by each agent, with periodic communication to align solutions. It is well-suited for multi-robot systems because it can handle constraints and converges quickly for many practical problems.
Why not just use reactive collision avoidance alone? Pure reactive methods (like potential fields or CBF-only control) avoid collisions but struggle with deadlocks—robots can get stuck oscillating near goals because they have no look-ahead coordination. Predictive planning (ADMM) avoids these deadlocks but may be too slow to guarantee safety if it doesn't converge fully.
Can this framework be applied to aerial or ground robots? Yes, the same hybrid ADMM-CBF architecture is robot-agnostic—it only requires a kinematic model and the ability to communicate with neighbors. It has been demonstrated on robotic boats, but could be adapted for drones, ground rovers, or even humanoid robots in tightly coordinated tasks.
Conclusion
This distributed motion planning framework successfully combines predictive ADMM planning with reactive CBF safety filters to enable self-reconfiguring robotic boats to form arbitrary shapes without collisions or deadlocks. The approach runs at 5 Hz on onboard computers and scales linearly with swarm size, marking a practical step toward autonomous aquatic construction teams.