Predicting trajectories of peripheral agents for safe mobile robot navigation in industrial environments
Loading...
URL
Journal Title
Journal ISSN
Volume Title
School of Engineering |
Master's thesis
Authors
Date
Department
Major/Subject
Mcode
Degree programme
Language
en
Pages
66
Series
Abstract
This thesis investigates short-horizon trajectory prediction of surrounding agents (humans and robots) as a proposed enabler for safer, more efficient control of Autonomous Mobile Robots (AMRs) in industrial environments. Conducted at Nokia Bell Labs, the work leverages the company’s Autonomous Robot Orchestration Solution, which provides factory-wide situational awareness, enabling trajectory predictions to be fed into the controller for collision anticipation. A reproducible pipeline is developed and evaluated to produce explainable predictions for the controller, suitable for future certification. The pipeline includes video-based detection and homography transformation, data filtering, deterministic predictors (Constant Velocity; Constant Turn-Rate and Velocity), and a proposed hybrid physics-aware method under development that seeks to balance interpretability and adaptability. Models are assessed on public pedestrian datasets and on in-house recordings of human and robot motion using Average Displacement Error (ADE) and Final Displacement Error (FDE) metrics. Results are reported as distributions; the acceptable accuracy range ultimately depends on safety margins to be specified in relevant safety standards and controller requirements. Within this evaluation, deterministic baselines provide strong short-horizon reference points. The proposed hybrid method has been iteratively improved and currently shows advantages over the baselines on the robot-motion dataset, whereas performance on pedestrian datasets and particularly on long-horizon FDE remains challenging. Overall, the pipeline is intended for integration with Nokia Bell Labs’ orchestrator to enable anticipatory safety in multi-agent, dynamic industrial environments.Description
Supervisor
Kyrki, VilleThesis advisor
David, PierreKhalili, Mata