Continuous-Time Trajectory Estimation for Asynchronous Robotic Sensing

This direction is merged into another direction: continuous-time trajectory modeling. Its last version stays readable and citable here.

Version 1 of 131 papers45 conceptsfrozen 2026-10-04

In classical robotics, we often treat a robot's path as a sequence of discrete "snapshots"—a list of poses at specific timestamps. While this discrete-time approach works for low-speed navigation with synchronized cameras, it breaks down when sensors operate at different rates or when a robot moves fast enough to distort its own measurements.

Modern state estimation has shifted toward continuous-time trajectory modeling. Instead of estimating a finite set of points, researchers model the robot's motion as a smooth, continuous function $x(t)$. This shift allows us to query the robot's state at any arbitrary microsecond, naturally resolving the "asynchronicity" that plagues multi-sensor systems.

The Problem with Discrete Snapshots

Most robotic sensors do not fire at the same time. An Inertial Measurement Unit (IMU) might provide data at 1000Hz, while a LiDAR scans at 10Hz and a camera captures frames at 30Hz. In a discrete-time framework, we are forced to "bin" these measurements into the nearest time step or use crude linear interpolation, which introduces errors.

Furthermore, scanning sensors like LiDAR do not capture a whole scene instantly. As Angela P. Schoellig, Keenan Burnett, and Timothy D. Barfoot point out in IMU as an input versus a measurement of the state in inertial-aided state estimation, a LiDAR point cloud is actually a "rolling" collection of points, each with its own timestamp. If the robot moves during the scan, the resulting map becomes warped. A continuous-time model solves this by allowing the system to query the exact pose for every individual laser pulse, effectively "un-distorting" the sensor data.

Two Paths to Continuity: Splines and Gaussian Processes

To represent a trajectory as a continuous function, the field has converged on two primary mathematical tools: B-splines and Gaussian Processes (GPs).

1. B-Splines: Local Control and Efficiency

B-splines represent a trajectory as a weighted sum of basis functions. They are popular because they offer "local control"—changing a control point only affects a small segment of the trajectory.

Elias Mueggler and Davide Scaramuzza demonstrated the power of this approach in Continuous-Time Visual-Inertial Trajectory Estimation with Event Cameras. Event cameras produce a nearly continuous stream of data; by using cubic splines, they could fuse these microsecond-resolution events with IMU data without the computational explosion of discrete-time states.

Recent advances have made splines even more efficient. Daniel Cremers and Vladyslav Usenko developed recursive formulations for B-splines on Lie groups, reducing the complexity of computing derivatives from cubic to linear (Efficient Derivative Computation for Cumulative B-Splines on Lie Groups). This allows real-time optimization even for complex, high-degree-of-freedom motion.

2. Gaussian Processes: Probabilistic Smoothness

Gaussian Processes offer a different intuition: they treat the trajectory as a realization of a stochastic process. In Robotics: Science and Systems X (2014), Timothy D. Barfoot and colleagues showed that if we model motion using linear stochastic differential equations (SDEs), the resulting GP has an "exactly sparse" inverse kernel.

This is a breakthrough for scalability. Normally, GPs are computationally expensive ($O(N^3)$), but this sparsity allows for trajectory estimation that scales linearly with the number of measurements. This framework has been extended to handle multi-sensor setups, such as the Traj-LIO system by Jianke Zhu and Xin Zheng, which remains resilient even when an IMU fails by relying on the GP's motion prior.

Rethinking the IMU: From Input to Measurement

A subtle but profound shift in this research direction involves how we treat inertial data. In discrete-time systems, the IMU is usually treated as a "control input" that propagates the state forward.

In the continuous-time paradigm, researchers like Angela P. Schoellig argue for treating the IMU as a measurement of the state's derivatives (velocity and acceleration). By modeling the trajectory as a GP, IMU readings become unary constraints on the curve's shape. This approach, seen in Continuous-Time Radar-Inertial and Lidar-Inertial Odometry using a Gaussian Process Motion Prior, allows for a cleaner separation of sensor noise and process noise, leading to significantly more robust estimation during aggressive maneuvers.

Why It Works: The Intuition of Smoothness

The core reason continuous-time modeling outperforms discrete methods is that it encodes physical priors directly into the representation. A robot cannot teleport; its position, velocity, and acceleration must change smoothly.

By using B-splines or GP motion priors, we enforce this smoothness mathematically. When a sensor provides a noisy or asynchronous measurement, the continuous-time framework doesn't just ask "where is the robot now?" but "what is the smoothest path that explains all these asynchronous observations?"

This perspective is now being applied to even more exotic domains, such as Neural Implicit Representations for camera poses and 3D Gaussian Splatting SLAM. As robots move faster and sensors become more specialized, the "snapshot" era is giving way to a world where motion is truly continuous.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_continuous_time_trajectory_estimation_fo_0bd627b3,
  title = {Continuous-Time Trajectory Estimation for Asynchronous Robotic Sensing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-time-trajectory-estimation-for-asynchronous-robotic-sensing/rdr_0bd627b386d34e419496592b1c356465},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_continuous_time_trajectory_estimation_fo_0bd627b3_v1,
  title = {Continuous-Time Trajectory Estimation for Asynchronous Robotic Sensing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-time-trajectory-estimation-for-asynchronous-robotic-sensing/rdr_0bd627b386d34e419496592b1c356465/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}