Abstract
Trajectory planning for tractor-trailer robots in unstructured environments is challenging due to complex kinematics, high-dimensional state spaces, and the risk of jackknifing. This paper presents Tracailer, an optimization-based planner that achieves several-fold efficiency improvements over state-of-the-art methods. By introducing a "slackened arc length" trajectory representation and leveraging direct continuous-space deformation for collision avoidance via Signed Distance Fields (SDF), Tracailer generates safe, high-order smooth paths. Experimental results in both simulation and real-world deployments demonstrate a significant reduction in planning time—from 4.14 seconds to 0.364 seconds—while maintaining high robustness to perception noise.
Introduction
Navigating a tractor-trailer robot presents unique motion planning challenges. Unlike standard vehicles, these robots possess a deformable structure where trailers pivot on hinges, leading to three primary difficulties: complex non-holonomic kinematics, a high-dimensional state space that grows with each added trailer, and computationally expensive collision checking. Traditional methods often rely on Safe Travel Corridors (STCs), which can be slow to compute and overly restrictive.
Tracailer addresses these issues by providing a lightweight, robust optimization framework capable of handling complex maneuvers, such as three-point turns, without the numerical instabilities typical of traditional differential flatness models.
Methodology
The Tracailer architecture is built upon three core innovations:
- Slackened Arc Length Representation: To avoid singularities when the robot stops or moves slowly, Tracailer introduces a variable $s(t)$ that decouples the spatial path from temporal timing. This "slackening" ensures high-order smoothness using quintic piecewise polynomials and stabilizes the optimizer during low-velocity maneuvers.
- Direct Trajectory Deformation: Instead of pre-calculating STCs, the planner utilizes a Signed Distance Field (SDF). The SDF provides a gradient that directly "pushes" the trajectory away from obstacles in continuous space, preserving the full solution space and reducing dependence on initial guesses.
- SE(2)-MHA (Multi-Terminal Hybrid-A*): A fast initial path search is performed in the SE(2) space of the tractor, using a quick approximation for the trailers. This provides a high-quality initial guess nearly an order of magnitude faster than full state-space searches.
Experimental Evaluation
The system was evaluated through simulations and real-world hardware tests. Simulations included "Parking Lot" and "Village" scenarios with narrow passages and random obstacles. Real-world tests utilized an AgileX Hunter SE tractor with three trailers and a large-scale industrial setup. Tracailer was benchmarked against OCP-STC, a standard optimization-based planner.
Results and Discussion
Tracailer demonstrated superior performance across several metrics:
- Efficiency: Total planning time was reduced by over 90%, from 4.14 seconds to 0.364 seconds. The SE(2)-MHA path finder was approximately 7x faster than standard Hybrid-A*.
- Path Quality: The generated trajectories exhibited lower average curvature and reduced tracking errors by nearly 20% compared to methods not using the slackened arc length.
- Robustness: The system maintained an 84% success rate under high perception uncertainty by replanning at high frequencies to correct for trailer-yaw estimation noise.
Conclusion
Tracailer offers a highly efficient and robust solution for tractor-trailer robot navigation in complex environments. While the current system excels in unstructured settings, future work will focus on improving performance in extremely dynamic environments and integrating semantic awareness to better predict the movement of specific obstacles.
The source code is open-source and available at https://github.com/Tracailer/Tracailer.
Summary updated