Self-Supervised Path Planning in Unstructured Environments via Global-Guided Differentiable Hard Constraint Projection

TL;DR

This work tackles safe, data-efficient path planning for embedded mechatronics in unstructured environments by introducing a self-supervised framework that fuses a Global-Guided Artificial Potential Field (G-APF) with a differentiable hard constraint projection layer. The approach uses slack-variable reformulations and an LSE-smoothed collision representation to enforce kinematic and collision constraints within an end-to-end trainable network, augmented by a two-stage curriculum training strategy. Empirical results on 20,000 static scenarios and CARLA-based closed-loop tests on NVIDIA Jetson Orin NX show an 88.75% planning success rate with a real-time latency of 94 ms, outperforming several baselines in safety and dynamic feasibility. The framework provides a generalizable pathway for embedding physical laws into neural planners, enabling reliable, real-time constrained optimization for mechatronics applications.

Abstract

Deploying deep learning agents for autonomous navigation in unstructured environments faces critical challenges regarding safety, data scarcity, and limited computational resources. Traditional solvers often suffer from high latency, while emerging learning-based approaches struggle to ensure deterministic feasibility. To bridge the gap from embodied to embedded intelligence, we propose a self-supervised framework incorporating a differentiable hard constraint projection layer for runtime assurance. To mitigate data scarcity, we construct a Global-Guided Artificial Potential Field (G-APF), which provides dense supervision signals without manual labeling. To enforce actuator limitations and geometric constraints efficiently, we employ an adaptive neural projection layer, which iteratively rectifies the coarse network output onto the feasible manifold. Extensive benchmarks on a test set of 20,000 scenarios demonstrate an 88.75\% success rate, substantiating the enhanced operational safety. Closed-loop experiments in CARLA further validate the physical realizability of the planned paths under dynamic constraints. Furthermore, deployment verification on an NVIDIA Jetson Orin NX confirms an inference latency of 94 ms, showing real-time feasibility on resource-constrained embedded hardware. This framework offers a generalized paradigm for embedding physical laws into neural architectures, providing a viable direction for solving constrained optimization in mechatronics. Source code is available at: https://github.com/wzq-13/SSHC.git.

Figures (7)

• Figure 1: Overall framework of the proposed self-supervised planning system. The architecture integrates a Global-Guided Artificial Potential Field to provide dense supervision signals for the Policy Network during the training phase. To ensure safety, an adaptive-depth neural projection (AdaNP) is embedded in the loop, which rectifies the raw network output into a feasible path by solving the constraint equation $\mathbf{g}(\mathbf{p}) + \mathbf{s} \odot \mathbf{s} = \mathbf{0}$.
• Figure 2: Visualization of the Global-Guided Potential Field. The multi-source wavefront propagation creates a convex valley along the optimal path, providing dense gradient guidance to steer the agent out of local minima.
• Figure 3: Global-Guided Potential Field Loss. (a) Direct potential minimization guides the agent toward lower potential regions. (b) Local attraction ensures continuous progress along the potential valley, especially in obstacle-dense areas.
• Figure 4: The evolution of the planned path during training. (a)-(c) Stage 1: The network learns to follow the potential valley with soft constraints roughly. (d)-(f) Stage 2: The hard projection layer refines the path to enhance safety and kinematic feasibility.
• Figure 5: Qualitative comparison of generated paths in two narrow-corridor scenarios (top and bottom rows). As highlighted in the zoomed-in regions: the NMPC baseline exhibits sparse waypoint distribution, which poses a risk of collision during actual execution due to interpolation errors. The comparison between the Soft and Hard methods demonstrates the effectiveness of the proposed hard constraint formulation in strictly guaranteeing collision-free swept areas within tight spaces.