# FlexPath: Adapting Learned Connectivity Guidance to Path Preferences

arXiv: https://arxiv.org/abs/2606.10167 · HTML: https://mlsanity.com/paper/2606.10167/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 19 of its 43 references that are in the map.
- First posted on arXiv 8 Jun 2026.

## Builds on (most cited first, 19 of 19)

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [Deep Residual Learning for Image Recognition](https://mlsanity.com/paper/1512.03385/) (2015)
- [U-Net: Convolutional Networks for Biomedical Image Segmentation](https://mlsanity.com/paper/1505.04597/) (2015)
- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [SGDR: Stochastic Gradient Descent with Warm Restarts](https://mlsanity.com/paper/1608.03983/) (2016)
- [Differentiation of Blackbox Combinatorial Solvers](https://mlsanity.com/paper/1912.02175/) (2019)
- [Motion Planning and Control for Mobile Robot Navigation Using Machine Learning: a Survey](https://mlsanity.com/paper/2011.13112/) (2020)
- [Path Planning using Neural A* Search](https://mlsanity.com/paper/2009.07476/) (2020)
- [Learning Heuristic Search via Imitation](https://mlsanity.com/paper/1707.03034/) (2017)
- [Theta*: Any-Angle Path Planning on Grids](https://mlsanity.com/paper/1401.3843/) (2014)
- [Learning Risk-Aware Costmaps via Inverse Reinforcement Learning for Off-Road Navigation](https://mlsanity.com/paper/2302.00134/) (2023)
- [TransPath: Learning Heuristics For Grid-Based Pathfinding via Transformers](https://mlsanity.com/paper/2212.11730/) (2022)
- [iA*: Imperative Learning-based A* Search for Path Planning](https://mlsanity.com/paper/2403.15870/) (2024)
- [An Autonomous Path Planning Method for Unmanned Aerial Vehicle based on A Tangent Intersection and Target Guidance Strategy](https://mlsanity.com/paper/2006.04103/) (2020)
- [The Emergence of Deep Reinforcement Learning for Path Planning](https://mlsanity.com/paper/2507.15469/) (2025)
- [Enhanced Robot Motion Block of A-star Algorithm for Robotic Path Planning](https://mlsanity.com/paper/2312.15738/) (2023)
- [Energy Optimized Robot Arm Path Planning using Differential Evolution in Dynamic Environment](https://mlsanity.com/paper/1806.08916/) (2018)

## Abstract (arXiv)

Recent learning-based path planners use neural networks to process occupancy representations and approximate heuristics for classical search algorithms, yielding near-optimal paths with reduced search effort. However, these methods are tied to a fixed objective, usually the shortest-path objective, implicit in their supervision. This limits their flexibility to accommodate alternative criteria. We introduce FlexPath, a two-stage learned search-guidance framework that first learns a recall-oriented connectivity prior initialized from shortest-path planner demonstrations and then refines this prior using differentiable path-shape objectives, thereby separating demonstration-based learning of connectivity-biased guidance from subsequent objective specific refinement. Beyond enabling adaptation to new routing preferences, the two-stage procedure improves standard shortest-path planning itself: on TMP, FlexPath improves optimal-path recovery from 75.0% to 88.6% over TransPath while reducing search expansions by 13.8%. Ablations show that neither prior learning nor objective fine-tuning alone matches the full pipeline; their combination yields the strongest path cost and search efficiency. We further demonstrate the preference adaptation by adapting guidance to non-shortest-path objectives such as obstacle clearance, class-conditioned obstacle clearance and waypoint following. For clearance with dmin=2, FlexPath achieves 96.2% full clearance satisfaction on feasible instances while maintaining low search effort, and it reaches 98.4% waypoint-following success.