FlowPed: Proprioceptive State Fusion and Efficient Diffusion for Pedestrian Trajectory Prediction in Manufacturing Plants

Autonomous Guided Vehicles (AGVs) must anticipate human movement to operate safely in manufacturing plants. Unlike outdoor crowd settings, shop floors offer rich, reliable environmental priors (CAD/HD maps, stations, sidewalks, flow rules), but little shareable trajectory data and tight latency budgets. We presentFlowPed, a trajectory forecasting framework tailored to this regime. FlowPed (i) fuses training-time privileged plant signals, including finite-automaton (FAM) motion states and geometry-aware features to stabilize learning under data scarcity; (ii) abstracts behavior via a VQ-VAE that discovers discrete latent “actions” and augments a gated residual + Transformer encoder to form a compact condition vector; and (iii) generates futures with an Elucidated Diffusion Model (EDM) implemented as a Diffusion Transformer (DiT) with few-step Karras sampling for real-time inference. We pretrained on public pedestrian repositories using only generic kinematics and fine-tuned on a VR shop-floor dataset (30 participants; 10 hours) instrumented with worker pose/gaze and AGV state. In ablations, FlowPed achieved 2.5× faster inference than a 5-step flow-matching baseline on AGV-class hardware (0.022 s vs. 0.056 s per forecast on an RTX A2000) while maintaining or improving accuracy (e.g., 0.6%$L\_{\text{2 mean}}$, 25.6%$L\_{\text{2 max}}$at 4s). Pre-training yielded consistent gains across decoder variants (4.2–6.4%$L\_{\text{2 mean}}$, 3.7–5.0%$L\_{\text{2 max}}$). The results indicate that the combination of plant priors, discrete latent actions, and few-step diffusion enables accurate, low-latency pedestrian prediction suitable for embedded AGV stacks in environment-rich, but data-scarce factories.