Evaluating Predictive World-Model Representations for Long-Horizon Forklift–Worker Risk Forecasting

Predictive world models can simulate plausible future physical scenes, but their value as frozen representations for specialized industrial safety tasks remains unclear. We examine whether predictive future latents from NVIDIA Cosmos Predict2.5 2B add sample-specific information beyond a compact kinematic history for long-horizon forklift–worker risk forecasting. From a 2 s state history, the task predicts both agents’ positions, their cumulative minimum separation, and HIGH/WARNING/SAFE risk from +1 to +6 s. Using 1180 samples from 854 synthetic warehouse runs, a state-only predictor is compared with a residual fusion of aligned Cosmos latents, the same residual with shuffled latents, and a Cosmos-only model. On 300 test samples from 129 runs, the state-only model achieves a worker average displacement error of 0.484 m and a mean risk Macro-F1 of 0.846; the aligned residual yields 0.496 m and 0.845, and the shuffled control 0.485 m and 0.850. Run-level cluster bootstrap analysis and ten strict shuffled controls show no consistent alignment-specific benefit, and this conclusion holds under alternative risk thresholds and spatially preserving latent probes. The latent contains weak scene information that is largely redundant with the metric state. Matched state baselines and sample-correspondence controls are essential for evaluating foundation-model representations for physical-safety tasks.