Source-linked AI summary
Safe Trajectory Generation for Complex Urban Environments Using Spatio-temporal Semantic Corridor
Wenchao Ding, Lu Zhang, Jing Chen, Shaojie Shen
TL;DR
Urban trajectory planning must accommodate heterogeneous semantic elements and avoid safety gaps between sampled points. The paper encodes semantics in connected spatio-temporal cubes and optimizes a trajectory with a quadratic program, reporting analysis across varied traffic configurations and semantic elements.
Problem
Urban AV planning must reconcile semantic elements represented as obstacles, constraints, or costs and address safety gaps that sampled-point checks may miss.
Source: §I · p. 1Method
The method represents semantics as connected collision-free cubes with dynamical constraints, then uses a piecewise Bezier-curve quadratic program for trajectory generation.
Source: §I · pp. 1–2Results
The authors report analysis across varied traffic configurations and complex semantic elements, including a construction scenario combining road layout, dynamic obstacles, and speed limits.
Source: 1) Merging into congested traffic due… · p. 7, §VIII · p. 8Takeaways & Limitations
A shared semantic abstraction targets the challenge of handling diverse element types and combinations within trajectory planning.
Source: §I · pp. 1–2Prediction uncertainty and interaction uncertainty are not sufficiently modeled.
Source: §VIII · p. 8
Abstract
from arXiv · showhide
Planning safe trajectories for autonomous vehicles in complex urban environments is challenging since there are numerous semantic elements (such as dynamic agents, traffic lights and speed limits) to consider. These semantic elements may have different mathematical descriptions such as obstacle, constraint and cost. It is non-trivial to tune the effects from different combinations of semantic elements for a stable and generalizable behavior. In this paper, we propose a novel unified spatio-temporal semantic corridor (SSC) structure, which provides a level of abstraction for different types of semantic elements. The SSC consists of a series of mutually connected collision-free cubes with dynamical constraints posed by the semantic elements in the spatio-temporal domain. The trajectory generation problem then boils down to a general quadratic programming (QP) formulation. Thanks to the unified SSC representation, our framework can generalize to any combination of semantic elements. Moreover, our formulation provides a theoretical guarantee that the entire trajectory is safe and constraint-satisfied, by using the convex hull and hodograph properties of piecewise Bezier curve parameterization. We also release the code of our method to accommodate benchmarking.
I. INTRODUCTION
Urban trajectory planning must reconcile heterogeneous semantic elements and guarantee safety across the entire trajectory, not only at sampled points. The paper proposes a unified spatio-temporal semantic corridor (SSC) and a trajectory-optimization formulation, building on prior semantic and corridor-planning approaches.
I. INTRODUCTION
- Urban planning must reconcile elements represented as obstacles, constraints, or costs, while sampled-point planners can miss infeasible motion between samples.The paper identifies both semantic heterogeneity and whole-trajectory safety as planning challenges.Source: §I · p. 1
- The SSC represents different semantic elements as connected collision-free cubes carrying their associated dynamical constraints.The paper describes the SSC as an abstraction that allows semantic elements to be combined without changing the planning formulation.Source: §I · pp. 1–2, A. Spatio-temporal motion planning for… · p. 2
- The trajectory is optimized within the SSC using a quadratic program and piecewise Bezier curves to guarantee safety and feasibility for the entire trajectory.The formulation relies on the convex-hull and hodograph properties of Bezier curves.Source: §I · p. 2
A. Spatio-temporal motion planning for AVs
- Prior spatio-temporal lattices omit some traffic semantics, while semantic cost fields can face local minima and balancing challenges.Other approaches represent traffic lights and agents as obstacles, or forbidden lane changes and speed limits as obstacles and constraints.Source: A. Spatio-temporal motion planning for… · p. 2
- The SSC builds on obstacle-like and constraint-like representations by abstracting semantic elements so new combinations do not require changes to cost formulation or constraint specification.The authors present this abstraction as the basis for a unified, generalizable trajectory-generation framework.Source: A. Spatio-temporal motion planning for… · p. 2
B. Corridor generation for AVs
The paper extends spatial corridor planning into the spatio-temporal domain to handle dynamic obstacles and semantic elements. Its system uses behavior-planner simulations to seed corridor generation and can prepare trajectories for multiple potential behaviors.
B. Corridor generation for AVs
- Earlier spatial corridors account for free space or vehicle dynamics but are static, while the proposed corridor extends planning to dynamic obstacles and semantic elements.A cited convex feasible-set approach is also described as too computationally complex for real-time applications and does not consider semantic elements.Source: B. Corridor generation for AVs · p. 2
III. SYSTEM OVERVIEW
- A planning cycle proceeds from environment understanding and prediction through behavior planning to motion planning, which uses behavior-planner states as corridor-generation seeds.The trajectory framework belongs to the motion-planning layer and can also work with other behavior planners that provide preliminary future-state guesses.Source: §III · pp. 2–3
- SSC construction uses a semantic map, predicted agent trajectories, forward-simulated states, and a reference lane; prediction can be optional when simulations already include other agents.The source of the seeds and the interaction modeled depend on the behavior planner.Source: §III · p. 3
IV. PRELIMINARIES ON MULTI-POLICY DECISION MAKING
- MPDM simplifies multi-agent behavior planning by simulating a finite set of closed-loop policies and selecting the best behavior with a reward function.Its forward-simulated ego states serve as seeds, but their coarse resolution and simplified simulation model make them unsuitable for direct execution.Source: §IV · p. 3
- Because MPDM supplies states for multiple behaviors, the framework generates candidate trajectories for each and can prepare a switch-back trajectory while executing a lane change.The paper illustrates this with lane-change-left, lane-change-right, and lane-keeping behaviors.Source: §IV · p. 3
A. Semantic Elements And Fren´et Frame Representation
The paper represents lane-related semantics in longitudinal, lateral, and time coordinates, mapping forbidden regions to obstacles and rule-like conditions to semantic boundaries. This representation supports generating connected corridor cubes around seeds while preserving high-level behavior.
A. Semantic Elements And Fren´et Frame Representation
- The Frenét frame uses longitudinal s and lateral l relative to a reference lane, with time as a third dimension for time-indexed semantics.The paper motivates these coordinates by lane-associated rules and time-profiled predicted trajectories.Source: A. Semantic Elements And Fren´et Frame… · p. 3
1) Obstacle-like semantic elements:
- Obstacle-like elements occupy forbidden regions in s-l-t, while constraint-like elements act over bounded spatial or temporal ranges.Static and dynamic obstacles, and red lights, are rendered as occupancy; speed limits and lane-change durations are represented by semantic boundaries marking where constraints apply.Source: 1) Obstacle-like semantic elements · p. 4, 2) Constraint-like semantic elements · p. 4
2) Constraint-like semantic elements:
- Traffic-rule constraints such as speed limits are hard, whereas lane-change-duration constraints are soft and handled during corridor generation.The paper states that soft constraints lack a universal quantitative description.Source: 2) Constraint-like semantic elements · p. 4
B. Semantic Corridor Generation
- Corridor generation uses seeds from forward-simulated states and an initial state, with consecutive-seed initial cubes required to be collision-free.The authors describe the resulting corridor as preserving the seeds’ high-level behavior; Fig. 4 illustrates inflation against a speed-limit boundary and obstacle, followed by relaxation.Source: 1) Seed Generation · p. 4, B. Semantic Corridor Generation · p. 4, Figure 4 · p. 4
1) Seed Generation:
Cube inflation builds connected corridor cubes from consecutive seeds and accounts for semantic boundaries as well as obstacles. Its stopping and direction rules align the cubes with constraints while preserving the piecewise trajectory structure.
1) Seed Generation:
- Consecutive seeds are treated as two cube vertices to construct an initial cube, while seeds already inside the last inflated cube are skipped as topologically equivalent.The process iterates over seeds and only constructs an initial cube when the seed is not contained in the last cube.Source: 2) Cube Inflation with Semantic… · pp. 4–5
- Inflation alternates among the three s-l-t directions and stops when a step collides with an obstacle or intersects a semantic boundary.When an initial cube intersects a boundary, inflation opposite the entry direction is disabled so the inflated cube can nearly match that boundary.Source: 2) Cube Inflation with Semantic… · p. 5
- Each cube corresponds to one trajectory piece, so adjacent cubes share a time bound to preserve convexity without optimizing piece durations.The authors note that optimizing durations is non-convex; additional overlap in time could help if durations were optimized.Source: 2) Cube Inflation with Semantic… · p. 5
3) Cube Relaxation:
Cube relaxation creates extra room for optimization while preserving hard constraints and collision-free space. For trajectory generation, piecewise Bézier curves address limitations of monomial trajectories and use convex-hull and hodograph properties.
3) Cube Relaxation:
- Cube relaxation leaves extra optimization space for soft constraints while preserving hard constraints and collision-free space.The allowed margin is determined by constraints on consecutive cubes, such as velocity-matching distance longitudinally or allowed lane-change-duration fluctuation laterally.Source: 3) Cube Relaxation · p. 5
VI. TRAJECTORY GENERATION WITH SAFETY AND FEASIBILITY GUARANTEE
- Piecewise Bézier curves address monomial trajectories’ limited maneuver representation and sampled-point checks that can miss collisions between samples.The paper uses the curves’ convex-hull and hodograph properties to remove these limitations.Source: §VI · p. 5
A. B´ezier Basis and Its Properties
Bézier convex-hull and hodograph properties let the planner constrain a curve and its derivatives through their control points. The trajectory is piecewise, with each segment associated with an SSC cube and a jerk-based cost that remains invariant across semantic-element combinations.
A. B´ezier Basis and Its Properties
- A Bézier curve stays within the convex hull of its control points, and the hodograph property lets derivative control points constrain its dynamical profile.Constraining curve control points to convex free-space guarantees a collision-free curve; applying the convex-hull property to the derivative curve confines dynamics to a specified range.Source: A. B´ezier Basis and Its Properties · p. 5
B. Piecewise B´ezier Curve Representation
- Each piecewise Bézier segment corresponds to one SSC cube, and its duration is handled by a segment-specific time scaling factor.The segment cost integrates squared jerk; the objective is described as simple and invariant across semantic-element combinations.Source: B. Piecewise B´ezier Curve… · pp. 5–6
C. Enforcing Safety and Dynamical Constraints
The method uses a quintic piecewise Bézier trajectory parameterization and its hodograph and convex-hull properties to impose derivative bounds across each segment.
C. Enforcing Safety and Dynamical Constraints
- The hodograph property gives boundary derivatives of the Bézier trajectory from its control points.These endpoint relations support imposing derivative constraints on the trajectory segments.Source: C. Enforcing Safety and Dynamical… · p. 6
- The convex-hull property lets the method constrain the entire derivative profile using sufficient bounds.The bounds are specified separately for each dimension and derivative order.Source: C. Enforcing Safety and Dynamical… · p. 6
1) Desired state constraints:
The trajectory is constrained to match the given initial and desired terminal states through linear equalities on its Bézier parameters.
1) Desired state constraints:
- The first and last trajectory segments are constrained to match the given endpoint derivatives for k = 0, 1, and 2.Applying Eq. 5 expresses these conditions as linear equality constraints on p.Source: 1) Desired state constraints · p. 6
2) Continuity constraints:
The method enforces trajectory continuity, collision-free motion within corridor cubes, and semantic and dynamical limits using linear constraints assembled into a QP.
2) Continuity constraints:
- Adjacent segments are constrained to be continuous through derivatives of orders k = 0, 1, 2, and 3.These continuity conditions become linear equality constraints on p.Source: 2) Continuity constraints · p. 6
3) Free-space constraints:
- Each trajectory segment is kept within its corresponding cube to ensure collision-free motion.The cube shape supplies position bounds for each dimension.Source: 3) Free-space constraints · p. 6
4) Dynamical constraints:
- Semantic and dynamical limits constrain first- and second-order derivatives, and the combined linear constraints form an efficiently solvable QP.The bounds correspond to maximum lateral and longitudinal velocity and acceleration; if no feasible solution exists, an error is returned to the behavior layer.Source: 4) Dynamical constraints · p. 6
A. Implementation Details
The framework is evaluated in simulation on three representative traffic cases and a benchmark track, including tests of semantic handling and collision avoidance.
A. Implementation Details
- The C++11 implementation runs stably at 20 Hz on an Intel I7-8700K desktop in the multi-agent simulation platform.The simulation includes interacting agents with unknown routes and limited ego-vehicle sensing.Source: A. Implementation Details · pp. 6–7
1) Merging into congested traffic due to road construction:
- In road construction, lane-change, dynamic-obstacle, and speed-limit scenarios, the SSC encodes information for optimization without requiring explicit semantic-type handling.The case tests these semantic elements together while merging into congested traffic.Source: 1) Merging into congested traffic due… · p. 7
2) Overtaking on an urban expressway:
- The urban-expressway case demonstrates safe, smooth overtaking at around 20 m/s, while prediction uncertainty remains insufficiently considered in SSC generation.The authors identify uncertainty modeling as future work.Source: 2) Overtaking on an urban expressway · p. 7
3) An unprotected left turn at an intersection:
- During an unprotected left turn, the method finds safe, feasible trajectories for interaction negotiation under a hard speed limit.The vehicle follows the behavior plan and navigates smoothly.Source: 3) An unprotected left turn at an… · p. 7
C. Comparisons and Analysis
- The comparison uses a satellite-map benchmark track with a red-light checkpoint, speed limit, and dense obstacles to test semantic response and collision avoidance.Both planners use MPDM behavior planning; the comparison is with the optimal-primitives method.Source: C. Comparisons and Analysis · p. 7
1) Collision-avoidance in cluttered environments:
The benchmark tests collision avoidance in clutter, precise stopping at a red light, and compliance with a low speed limit. The reported results favor the proposed method on arrival time and stopping behavior, while the conclusion notes that prediction and interaction uncertainty remain under-modeled.
1) Collision-avoidance in cluttered environments:
- Our method reaches the red light at 42 s, about 14 seconds earlier than, with smoother acceleration within the dynamical limit.In clutter, the benefit of the corridor representation is more apparent because many baseline primitives become infeasible; the paper attributes the comparison to continuous solution space versus discretization and limited state-space coverage.Source: 1) Collision-avoidance in cluttered… · pp. 7–8
2) Precise stop with a high entry speed:
- With a 13 m/s entry speed, our method achieves a precise stop while keeping maximum deceleration within the dynamic range.Unlike, it explicitly enforces the stopping boundary condition; the baseline requires a stopping mode to consistently reach the target boundary condition during replanning.Source: 2) Precise stop with a high entry speed · p. 8
3) Collision avoidance under a low speed limit:
- Our method strictly follows the 4 m/s speed limit; the paper also reports experiments with obstacles placed online.The online-obstacle experiments are described without a specific outcome in the supplied passage.Source: 3) Collision avoidance under a low… · p. 8
VIII. CONCLUSION AND FUTURE WORK
- The conclusion identifies insufficient modeling of prediction and interaction uncertainty as the main limitation and a current research direction.It also notes that Bézier curves may be useful for nonlinear trajectory optimization for autonomous vehicles.Source: §VIII · p. 8