Abstract
Multi-robot teams have attracted attention from industry and academia for their ability to perform collaborative tasks in unstructured environments, such as wilderness rescue and collaborative transportation. In this paper, we propose a trajectory planning method for a non-holonomic robotic team with collaboration in unstructured environments. For the adaptive state collaboration of a robot team to catch and transport targets to be rescued using a net, we model the process of catching the falling target with a net in a continuous and differentiable form. This enables the robot team to fully exploit the kinematic potential, thereby adaptively catching the target in an appropriate state. Furthermore, the size safety and topological safety of the net, resulting from the collaborative support of the robots, are guaranteed through geometric constraints. We integrate our algorithm on a car-like robot team and test it in simulations and real-world experiments to validate our performance. Our method is compared to state-of-the-art multi-vehicle trajectory planning methods, demonstrating significant performance in efficiency and trajectory quality.
Introduction
Multi-robot trajectory planning (MVTP) is a well-studied field, but tasks involving "catching and transporting" objects add layers of complexity that challenge traditional algorithms. In scenarios like wilderness rescue, robots must act as a single, flexible organism while navigating cluttered environments.
This task presents two primary challenges. First, kinematic constraints: car-like robots are non-holonomic, meaning they cannot move sideways and possess a limited turning radius. Second, strong coupling: the robots are physically linked by a net, necessitating specific safety guarantees. These include Size Safety, ensuring robots stay close enough to prevent the net from tearing, and Topological Safety, ensuring robots do not cross net edges and become entangled. Existing methods often struggle with the high dimensionality of these problems, leading to slow computation times or suboptimal "greedy" behavior.
Methodology
The proposed framework bridges the gap between precision and speed using several mathematical innovations:
- Differential Flatness: By representing the state of a car-like robot using "flat outputs" (the x,y coordinates of the rear axle), the complex control inputs are mapped to a 2D space, drastically reducing the optimization search space.
- Differentiable Catching Model: Rather than treating a catch as a binary event, the authors model it as a continuous, differentiable cost function. This allows the optimization algorithm to use gradient descent to deform the robot formation into the ideal shape to intercept a target.
- Geometric Safety Constraints: Using a Minkowski difference-based algorithm, the system calculates the "signed distance" between robots and obstacles. This ensures collision avoidance while maintaining the net's physical integrity and topology.
- Unconstrained Optimization: The entire constrained problem is reformulated into an unconstrained one using penalty functions, solved efficiently via the L-BFGS algorithm.
Experimental Results
The algorithm was evaluated through extensive simulations and real-world experiments. In simulations featuring up to 26 agents, the proposed method maintained a 100% success rate, while competing methods like Sequential Convex Programming (SCP) saw success rates plummet to 15% or lower at high agent counts.
Computationally, the method is highly efficient. For a 26-robot team, it required only 1.9 seconds to plan, compared to 376 seconds for the next best optimizer. In real-world tests involving three car-like robots catching a falling drone, the system generated rescue trajectories in just 0.053 seconds.
Discussion and Conclusion
A key insight of this work is adaptive catching: the robot team does not simply move to a point but actively deforms its formation to maximize the catchable area. By treating net edges as dynamic obstacles, the planner prevents entanglement without requiring complex cloth physics simulations.
While the current framework is centralized and operates in 2D, it demonstrates that reformulating complex robotic constraints into differentiable functions allows for real-time, high-dimensional collaborative planning. Future work will explore extending these principles to 3D environments and decentralized control schemes for larger swarms.
Summary updated