Autonomous vehicles and mobile robots must navigate environments defined by structured pathways, such as lanes, tracks, or pre-planned routes. However, standard onboard sensors and global positioning systems typically represent the world using a global Cartesian grid $(x, y)$. When planning trajectories or predicting the behavior of other agents in this absolute coordinate frame, even a simple task like following a curved road requires complex, non-linear trigonometric calculations.
To bypass this complexity, researchers use path-relative coordinate transformations (often formalized via the Frenet-Serret frame). Instead of tracking absolute positions, this technique maps the environment onto a reference path, such as a lane centerline. A vehicle's state is decomposed into two orthogonal components: longitudinal progress ($s$) along the path and lateral deviation ($d$) from it.
As noted by Lee et al. in Perceptual Attention-based Predictive Control, mapping trajectories directly to raw pixel or absolute coordinates without a structured mathematical representation of paths is highly inefficient. By embedding the underlying infrastructure's geometry directly into the state space, path-relative transformations simplify optimization, stabilize kinematic estimation, and enable models to generalize across diverse geographic layouts.
Decoupling Motion to Simplify Optimization and Control
The primary mathematical advantage of a path-relative frame is that it decouples longitudinal progress from lateral maneuvers. In a standard Cartesian frame, avoiding an obstacle requires simultaneous adjustments to both $x$ and $y$ coordinates, which are tightly coupled through the vehicle's heading angle.
Transforming this problem into a road-relative frame simplifies complex two-dimensional spatial constraints into independent one-dimensional adjustments. As shown by Liu et al. in Occupancy Prediction-Guided Neural Planner for Autonomous Driving, this coordinate reduction makes the calculation of safety, comfort, and progress costs significantly more computationally efficient for real-time motion planning.
Cartesian Frame (x, y) Frenet Frame (s, d)
(Highly Coupled) (Decoupled Motion)
___ |
/ \ <-- Curved Road | O (Obstacle)
/ O \ |-------
_________/_______\________ __________|_______|__________ Reference Path
Vehicle -> / Vehicle ->
/ |
This decoupling is particularly valuable for real-time control systems operating under tight computational budgets:
- High-Speed Racing: In FSDP: Fast and Safe Data-Driven Overtaking Trajectory Planning for Head-to-Head Autonomous Racing Competitions, Hu et al. formulate Model Predictive Control (MPC) optimization directly within the Frenet frame. This choice streamlines trajectory generation under dynamic constraints, making it amenable to real-time execution on resource-constrained hardware.
- Collision Avoidance: Schratter et al. demonstrate in Pedestrian Collision Avoidance System for Scenarios with Occlusions that control logic can be generalized across varying road geometries by separating movement into longitudinal and lateral components, simplifying collision risk calculations regardless of road curvature.
- Marine Robotics: The framework applies equally to marine environments. Dong et al. in Path Following Control of the Underactuated USV Based On the Improved Line-of-Sight Guidance Algorithm and Feng et al. in USV Application Scenario Expansion Based on Motion Control, Path Following and Velocity Planning utilize the Serret-Frenet coordinate system to define tracking errors relative to a moving point on a desired trajectory, simplifying error analysis and control design for underactuated surface vessels.
By abstracting the path geometry, planners can handle complex maneuvers fluidly. Han et al. show in Spatio-temporal Keyframe Control of Traffic Simulation using Coarse-to-Fine Optimization that simulating vehicle dynamics in Frenet coordinates allows agents to deviate from a central path to avoid collisions or meet timing constraints while still adhering to the general road topology. Similarly, Ding et al. combine Frenet frames with space-time diagrams in Energy-efficient Hybrid Model Predictive Trajectory Planning for Autonomous Electric Vehicles to resolve path geometry and obstacle avoidance within a unified framework.
Geographic Transferability and Invariant Learning
Deep learning models trained on absolute Cartesian coordinates frequently struggle with generalization. A neural network trained to predict trajectories or make tactical decisions using global coordinates often memorizes specific intersection layouts or road curvatures, causing it to fail when deployed in a new city.
Path-relative transformations overcome this limitation by removing absolute spatial dependencies. By mapping environmental data onto a curvilinear coordinate system aligned with lane paths, complex maneuvers are simplified into linear representations of progress and deviation. As argued by Zhang et al. in Map-Adaptive Goal-Based Trajectory Prediction, this abstraction prevents the system from needing to memorize specific intersection layouts, resulting in stable predictive performance across geographically distinct cities.
This spatial invariance benefits several facets of learning-based autonomous driving:
- Behavior Modeling: In Generic Prediction Architecture Considering both Rational and Irrational Driving Behaviors, Hu et al. demonstrate that road-aligned coordinates simplify the definition of rational driving constraints. Decoupling progress from deviation allows cost functions to easily penalize lane departures across complex geometries like roundabouts.
- Trajectory Prediction: Xin et al. show in Intention-aware Long Horizon Trajectory Prediction of Surrounding Vehicles using Dual LSTM Networks that representing vehicle motion as lateral deviation relative to target lane markings allows long-horizon predictors to remain accurate on curved paths without the computational overhead of high-resolution spatial grids. To enforce infrastructure constraints directly, Kisa introduces a model-agnostic framework in Stay on Track: A Frenet Wrapper to Overcome Off-road Trajectories in Vehicle Motion Prediction that transforms inputs and outputs into lane-relative coordinates, reducing off-road predictions by over 90%.
- Policy Generalization: For reinforcement learning, Kurzer et al. show in Generalizing Decision Making for Automated Driving with an Invariant Environment Representation using Deep Reinforcement Learning that mapping environmental data onto the agent's future trajectory achieves invariance to road curvature and intersection layouts. This ensures the policy focuses on the universal physics of gap acceptance and timing. Similarly, Wang et al. demonstrate in Learning hierarchical behavior and motion planning for autonomous driving that a lane-relative curvilinear frame ensures tactical maneuvers appear consistent whether the vehicle is on a straight or winding path.
This principle extends to broader spatiotemporal tasks. Landi demonstrates in Mobility Data Representations for Spatiotemporal Tasks that re-centering trajectory sub-sequences to a relative origin normalizes the data, allowing models to focus on geometric shape and movement dynamics rather than absolute coordinates. As further detailed by Landi et al. in Shape-based methods in mobility data analysis: effectiveness and limitations, applying shift normalization to subtrajectories removes absolute spatial context, enabling geographic transferability across entirely different urban environments.
Stabilizing Kinematics and Geometric Alignment
Representing motion relative to a path also stabilizes kinematic state estimation. In a standard Cartesian tracking system, estimating a vehicle's heading (yaw) along a curve is highly challenging because the absolute orientation angle constantly changes as the vehicle turns.
In Extended Object Tracking in Curvilinear Road Coordinates for Autonomous Driving, Dahal demonstrates that tracking vehicle states in a road-aligned curvilinear coordinate system stabilizes orientation estimation. Because the relative heading of a vehicle following a curve remains near zero, the complexity of the motion estimation task is reduced, directly providing the lane-relative information required for high-level driving tasks like lane-keeping or adaptive cruise control.
Cartesian Tracking (Yaw changes) Curvilinear Tracking (Relative Yaw ~ 0)
/ Yaw = 45° | Relative Yaw = 0°
/ |
/ |
| Yaw = 90° | Relative Yaw = 0°
| |
Path-relative structures also allow models to capture multi-agent interactions and environmental constraints more effectively:
- Dual-Coordinate Representations: Song et al. show in Learning to Predict Vehicle Trajectories with Model-based Planning that combining Cartesian and Frenet frames allows prediction models to capture lane-following behaviors while maintaining absolute kinematic constraints. For high-speed multi-vehicle racing, Liniger et al. demonstrate in Motion Planning and Control for Multi Vehicle Autonomous Racing at High Speeds that modeling dynamic agents via longitudinal speed and lateral offset within a Frenet frame yields a computationally efficient forecasting model suitable for extreme lateral accelerations.
- Structural Constraints: In Rethinking Lanes and Points in Complex Scenarios for Monocular 3D Lane Detection, Ye introduces lateral alignment constraints that relate points across different lanes sharing the same longitudinal coordinate. This architectural prior enables the model to better represent parallel structures and consistent curvature across the road surface.
- Target Alignment: To mitigate human driving noise during training, Huang et al. show in Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving that replacing expert demonstrations with perceived lane centerlines as the primary spatial targets forces the motion planner to prioritize the underlying road geometry over minor steering drifts.
Even when refining trajectories locally, path-relative framing provides better geometric context. Zhou et al. demonstrate in SmartRefine: A Scenario-Adaptive Refinement Framework for Efficient Motion Prediction that transforming environmental data into a local coordinate system centered on predicted future waypoints (anchors) allows models to better capture local geometric nuances along the intended path. For non-standard layouts like roundabouts, Cai shows in Topometric Imitation Learning For Route Following Under Appearance Change that employing 2D coordinate paths as local navigational commands provides the necessary geometric flexibility while remaining compatible with lightweight topological graphs.
Key Foundational Work
Generic Prediction Architecture Considering both Rational and Irrational Driving Behaviors (2019) — contributed 2 concept elements to this direction.
Intention-aware Long Horizon Trajectory Prediction of Surrounding Vehicles using Dual LSTM Networks (2019) — contributed 2 concept elements to this direction.
Pedestrian Collision Avoidance System for Scenarios with Occlusions (2019) — contributed 1 concept elements to this direction.