Technical Summary: FlexPath

Problem Statement

Learning-based path planners have recently utilized neural networks to approximate heuristics for classical search algorithms (e.g., A*), significantly reducing search effort while yielding near-optimal paths. However, existing methods are fundamentally constrained by their supervision signals, which are typically derived from shortest-path demonstrations. This ties the learned guidance to a fixed objective, limiting flexibility. Adapting these models to alternative criteria—such as obstacle clearance, class-conditioned avoidance, or waypoint following—usually requires retraining from scratch or manually designing complex cost functions, which is labor-intensive and struggles to generalize. Furthermore, current approaches often fail to separate the learning of path connectivity (structure) from specific path preferences (objectives), leading to suboptimal performance when objectives change.

Methodology: FlexPath

The authors introduce FlexPath, a two-stage learned search-guidance framework designed to decouple the learning of path connectivity from objective-specific refinement. The framework operates on a semantic grid input $X \in \mathbb{R}^{3 \times H \times W}$, where pixels encode semantic classes (free space, obstacles, start, goal, etc.).

Stage 1: Learning a Connectivity Prior

The first stage focuses on learning a recall-oriented connectivity prior. Instead of strictly imitating the exact shortest path, the model learns to predict connected traversable regions between start and goal.

• Imitation: The network is trained on rasterized shortest-path trajectories ($Y$) generated by a classical A* planner.
• Objective: A weighted binary cross-entropy loss is used, with a weight $\alpha > 0.5$ applied to false negatives. This biases the model toward high recall, ensuring the predicted path mask is slightly over-complete but topologically connected. This "over-complete" prior preserves connectivity, which is crucial for the subsequent refinement stage.
• Output: A soft path mask (Path Probability Map, PPM) $C \in [0, 1]^{H \times W}$ representing the confidence that a pixel lies on a feasible path.

Stage 2: Preference Adaptation via Path Shape Objectives (PSOs)

The second stage refines the connectivity prior $C$ to satisfy specific task objectives without relearning the underlying path structure. This stage uses differentiable Path Shape Objectives (PSOs) rather than new path labels.

• Mechanism: The soft path mask $C$ is optimized using a loss function composed of weighted PSOs. These functionals are differentiable with respect to $C$.
• Structural PSOs: Ensure geometric feasibility (collision avoidance and connectivity).
  • Connectivity: For shortest-path tasks, a soft-minimum cost propagation penalizes inefficient routing. For clearance tasks, a simpler reachability check (iterative expansion) is used.
  • Collision: Penalizes path confidence on obstacle cells.
• Task PSOs: Encode specific preferences.
  • Cost Minimization: Approximates discrete path cost via pixel-wise convolution.
  • Obstacle Clearance: Penalizes paths approaching obstacles closer than a threshold $d_{min}$ using a windowed log-sum-exp distance approximation.
  • Waypoint Following: Modifies connectivity constraints to enforce passage through intermediate waypoints.
• Adaptation: The same Stage 1 prior can be fine-tuned for different objectives (e.g., switching from shortest-path to clearance) by changing the PSO composition and loss weights, without retraining the full model or collecting new path labels.

Inference

At test time, the learned soft mask $C$ is converted into an inadmissible secondary heuristic ($h_C = \gamma(1 - C)$) to bias the search. A graph-based planner (Focal Search for shortest paths, Multi-Heuristic A* for other objectives) uses this heuristic alongside an admissible octile heuristic to extract a discrete trajectory.

Key Contributions

1. Imitation-to-Optimization Framework: The paper proposes a two-stage pipeline that first learns a reusable, recall-oriented connectivity prior via imitation learning and subsequently refines it using differentiable PSOs. This separates path-structure learning from objective-specific preferences.
2. Complementary Stages: The authors demonstrate that prior learning and objective fine-tuning are complementary. Neither stage alone achieves the performance of the full pipeline; their combination yields superior path cost, optimal-path recovery, and search efficiency.
3. Flexible Adaptation: The framework allows a single pretrained Stage 1 prior to be adapted to qualitatively different non-shortest-path objectives (obstacle clearance, class-conditioned avoidance, waypoint following) without retraining the full model or acquiring new pixel-level path labels.

Experimental Results

The authors evaluated FlexPath on the Tiled Motion Planning (TMP) dataset and zero-shot benchmarks (VoxelGym, City/Street Map, Starcraft).

• Shortest-Path Performance: On TMP, FlexPath improved optimal-path recovery from 75.0% (TransPath) to 88.6%, while reducing search expansions by 13.8%. It also achieved a higher "Hard Validity" (99.1% vs. 89.2%), indicating the raw PPM is more likely to form a valid connected path.
• Generalization: FlexPath consistently ranked first or second in cost and efficiency across zero-shot benchmarks (VG, CSM, SC), showing robustness to distribution shifts where baselines degraded.
• Preference Adaptation:
  • Obstacle Clearance ($d_{min}=2$): Achieved 96.2% full clearance satisfaction on feasible instances, outperforming classical cost-map shaped WA* (83.6%).
  • Class-Conditioned Clearance: Successfully distinguished obstacle classes, achieving 68.4% full clearance satisfaction for specific classes.
  • Waypoint Following: Reached 98.4% success in connecting start-to-waypoint-to-goal.
• Ablation Studies:
  • Removing the recall-biased weighting ($\alpha > 0.5$) in Stage 1 led to fragmented predictions and poor Hard Validity.
  • Fine-tuning from a random initialization (skipping Stage 1) resulted in model collapse for non-optimal objectives, confirming the necessity of the connectivity prior.
  • Binary occupancy inputs performed similarly to RGB inputs on TMP, suggesting the RGB encoding primarily aids flexibility for adding new semantic classes without retraining.

Significance and Claims

The paper claims that FlexPath demonstrates that learned search guidance need not remain bound to the objective represented by its training demonstrations. By separating the learning of path connectivity from objective-specific refinement, the framework achieves:

1. Improved Shortest-Path Planning: The two-stage process itself enhances standard shortest-path performance over established learned planners.
2. Efficient Adaptation: The ability to adapt to new routing preferences (safety, semantics, waypoints) via objective-level supervision rather than retraining from scratch.
3. Robustness: The learned guidance transfers effectively across unseen map domains and topologies.

The authors position this work as a step toward more flexible and generalizable path planning systems that can accommodate diverse real-world constraints without the prohibitive cost of retraining or manual cost-function engineering.