In conventional setups, when a person is lost from view, the robot typically stalls at the last seen position. IO-TEB instead lets the optimizer work with the predicted estimate, and when the person remains hidden for longer stretches, the robot actively searches along paths the person is likely to have taken, such as corridors and doorways. The researchers evaluated the planner across 4,320 simulated episodes spanning eight indoor scenes, three durations of occlusion, and real pedestrian trajectory data.

IO-TEB completed 87.36 percent of following tasks, compared with 55.14 percent for the standard timed elastic band planner. An important benchmark comparison involved a conventional design that handles the same probabilistic estimate in a separate finite-state supervisor, switching between modes depending on whether the person is visible. IO-TEB was statistically equivalent to that design, suggesting that embedding the belief estimate directly inside the optimizer costs neither accuracy nor computational overhead, while simplifying the architecture.

Real-time performance was another focus. The team compared IO-TEB with model predictive control using trajectory prediction and with three recent person-following planners. IO-TEB was not the most accurate of these methods, but in the authors' implementation it missed the 100-millisecond control-cycle deadline least often, exceeding the deadline on 3.32 percent of cycles, against a range of 5.14 to 44.21 percent for the alternatives.

Ablation experiments, in which components were removed one at a time, revealed which parts of the design actually drive the improvement. Replacing the last observation with the predicted estimate as the following reference accounted for most of the gain. By contrast, the additional terms that scale with the estimate's uncertainty brought no consistent benefit, a finding that could simplify future implementations.

The results are limited to simulation, albeit with realistic pedestrian trajectories, and the authors note the reported deadline-miss figures apply to their specific implementation. Even so, the study offers a concrete recipe for person-following robots: rather than adding external state machines or heavier predictors, a local planner can absorb a belief about a hidden target and remain computationally lean. The open-access study suggests that modest changes to existing planners, rather than wholesale redesigns, may deliver much of that robustness.

Subject of Research: Robot local planning for indoor person following under intermittent observations

Article Title: Belief-integrated timed elastic band planning for robust indoor person following under intermittent observations

Article References: Guan, J., Liu, S., Guan, K., Zheng, H., & Hu, W. (2026). Belief-integrated timed elastic band planning for robust indoor person following under intermittent observations. Scientific Reports. https://doi.org/10.1038/s41598-026-76768-3

Image Credits: AI Generated

DOI: 10.1038/s41598-026-76768-3

Keywords: robotics, person following, path planning, timed elastic band, probabilistic estimation, occlusion, trajectory optimization, model predictive control, indoor navigation, real-time planning, Belief-integrated, timed