Continuous Time Safety Verification for Trajectories

75 concepts55 papers1 proposal

When an autonomous system plans a path or verifies its safety, it typically reasons about discrete points in time or space. For instance, a model predictive controller might check for collisions every 10 milliseconds. However, physical systems operate in continuous time. If a drone flies at high speed, it could be perfectly safe at $t = 0.1\text{s}$ and $t = 0.2\text{s}$, but crash into a thin obstacle at $t = 0.15\text{s}$. This is the discretization gap. Traditional sampling-based collision checking or discrete-time verification leaves dangerous blind spots between nodes.

To achieve true safety, we need mechanisms that guarantee constraints hold across an entire continuous interval without requiring infinite computation. Two primary mathematical frameworks have emerged to bridge this gap: the convex hull property of parametric curves and Lipschitz continuity bounds.

The Convex Hull Property of Polynomial Trajectories

One of the most elegant ways to guarantee continuous-time safety is to represent trajectories using specific polynomial bases, such as Bernstein polynomials (Bézier curves) or B-splines. These representations possess a powerful geometric feature known as the convex hull property: the entire continuous curve is mathematically guaranteed to reside within the convex hull formed by its discrete control points.

Instead of checking every point along a continuous path $\mathbf{p}(t)$ for $t \in [0, T]$, we only need to ensure that the finite set of control points lies within a safe, convex geometric region (often called a safe corridor or polytope). If the control points are safe, the entire continuous trajectory is guaranteed to be safe.

This insight forms the backbone of modern high-speed trajectory planners. For example, James A Preiss et al. leverage Bézier curves to confine quadrotor trajectories within safe convex polytopes in Trajectory Optimization for Self-Calibration and Navigation. Similarly, Wenchao Ding et al. introduce a Spatio-temporal Semantic Corridor (SSC) in Safe Trajectory Generation for Complex Urban Environments Using Spatio-temporal Semantic Corridor, translating dynamic urban driving rules into safe space-time "cubes." By optimizing piecewise Bézier curves within these cubes, they eliminate the safety gaps inherent in discrete collision checking.

This framework scales efficiently to multi-agent systems. Jungwon Park et al. use piecewise Bernstein polynomials in Online Distributed Trajectory Planning for Quadrotor Swarm with Feasibility Guarantee using Linear Safe Corridor to enforce linear constraints on control points for distributed quadrotor swarms, preventing deadlocks. H. Jin Kim et al. combine this property with a Relative Safe Flight Corridor (RSFC) in Fast Trajectory Planning for Multiple Quadrotors using Relative Safe Flight Corridor to linearize inter-agent collision avoidance, transforming a non-convex problem into a high-speed Quadratic Program capable of coordinating dozens of drones.

Beyond quadrotors, the convex hull property simplifies complex constraints into basic algebraic operations. Yun-Woo Lee et al. demonstrate in BPMP-Tracker: A Versatile Aerial Target Tracker Using Bernstein Polynomial Motion Primitives that Bernstein basis functions allow autonomous systems to evaluate thousands of candidate paths in milliseconds for real-time target tracking. In the domain of fixed-wing aircraft, Luca Morando et al. exploit this property alongside differential flatness in Trajectory Planning and Control for Differentially Flat Fixed-Wing Aerial Systems to manage complex flight corridors as simple linear constraints on control points.

The same principle applies to B-splines. Boyu Zhou shows in Robust and Efficient Quadrotor Trajectory Generation for Fast Autonomous Flight that constraining B-spline control points transforms continuous safety and dynamic limits into a discrete set of checks, significantly reducing computational overhead while maintaining mathematical rigor. Wenchao Ding similarly utilizes this property in An Efficient B-spline-Based Kinodynamic Replanning Framework for Quadrotors to guarantee trajectory safety and feasibility by applying constraints directly to control points rather than the continuous curve itself. Furthermore, Xin Zhou et al. utilize uniform B-splines in EGO-Planner: An ESDF-free Gradient-based Local Planner for Quadrotors to build a local planner that guarantees safety simply by keeping control points collision-free.

Lipschitz Bounds for Continuous-Time Generalization

When trajectories cannot be easily parameterized as polynomials, or when verifying learned safety functions (such as neural network barrier certificates), researchers turn to Lipschitz continuity. A function is Lipschitz continuous if its rate of change is bounded by a constant $L$. If we know how fast a safety margin can change, we can evaluate safety at discrete points and mathematically bound the system's behavior between those points.

Suppose a safety function $h(x)$ must remain positive ($h(x) \geq 0$) for all states. If we verify that $h(x_i) \geq \epsilon$ at a discrete sample point $x_i$, and we know the function's Lipschitz constant is $L$, then within a radius of $\delta$, the function cannot drop below $h(x_i) - L\delta$. By choosing a "tightening parameter" $\epsilon = L\delta$, we guarantee that the function remains strictly positive everywhere between the samples.

This mechanism is widely used to extend discrete observations to continuous-time guarantees. S Roberts et al. utilize the Lipschitz continuity of a safety criterion function in Conservative collision prediction and avoidance for stochastic trajectories in continuous time and space to provide continuous-time collision avoidance, preventing agents from colliding between sampled time points. For data-driven verification, Luke Rickard et al. apply a tightening parameter based on Lipschitz constants in Continuous-time Data-driven Barrier Certificate Synthesis to verify continuous-time safety using only discrete-time trajectory observations.

This approach also underpins the safety of learning-based control systems. Weiye Zhao et al. establish safety certificates over continuous domains in Probabilistic Safeguard for Reinforcement Learning Using Safety Index Guided Gaussian Process Models by verifying safe control actions on a finite grid and using Lipschitz continuity to interpolate between them. Similarly, in the context of stability and reinforcement learning, Felix Berkenkamp et al. in Safe learning of regions of attraction for uncertain, nonlinear systems with Gaussian processes and Julia Vinogradska et al. in Stability of Controllers for Gaussian Process Forward Models use Lipschitz constants to guarantee that stability properties verified on a discrete grid generalize across the entire continuous state space.

Algebraic and Geometric Certificates

To further reduce the conservatism of global bounds, recent methods combine these geometric and analytical insights into localized certificates. In model predictive control, global convex hull approximations can sometimes be too restrictive, preventing agile maneuvers. To solve this, Jean Pierre Allamaa et al. partition the prediction horizon into smaller regions and constrain regional convex hulls in Real-time MPC with Control Barrier Functions for Autonomous Driving using Safety Enhanced Collocation, ensuring continuous-time safety between discretization nodes without excessive conservatism.

Alternatively, safety can be certified algebraically using separating hyperplanes. Alexandre Amice et al. certify collision avoidance between moving convex bodies in Certifying Bimanual RRT Motion Plans in a Second by finding a time-indexed hyperplane that separates them throughout the trajectory. By modeling the hyperplane's parameters as polynomials of time, they provide a continuous certificate of safety that eliminates the risk of missing collisions between sampled points.

By shifting the verification burden from infinite temporal sampling to finite geometric and algebraic constraints, these approaches allow autonomous systems to operate with absolute mathematical certainty in continuous physical space.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_convex_hull_safety_property_18505a51,
  title = {convex hull safety property},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/convex-hull-safety-property/rdr_18505a51b87e43d88a7b2c2a128237ac},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_convex_hull_safety_property_18505a51_v1,
  title = {convex hull safety property},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/convex-hull-safety-property/rdr_18505a51b87e43d88a7b2c2a128237ac/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake