Uncertainty-Aware Multi-Modal Learning for Autonomous Pallet Handling in Dynamic Warehouse Scenarios

Autonomous pallet transport in dynamic warehouses, where human workers and automated guided vehicles (AGVs) coexist, remains highly challenging due to sensory heterogeneity, environmental uncertainty, and strict real-time requirements. To address these issues, we propose an adaptive learning framework that integrates multi-modal sensory inputs, an Interactive Multiple Model (IMM) Kalman filter for trajectory forecasting, and Proximal Policy Optimization (PPO) for policy learning. A hybrid fusion strategy—combining early latent projection and shrinkage-enhanced cross-attention—ensures temporal-spatial alignment and suppresses noisy modalities, while a multi-objective reward balances safety, efficiency, and task success. Experiments in a high-fidelity warehouse simulator show that our method achieves a 78.5 % success rate, reducing collisions by 29 % and improving trajectory stability by 38% compared to PPO, MPC, SAC, and Transformer-based baselines. Real-time evaluation further confirms an average cycle latency of 33.3 ms, meeting industrial deployment standards. These results demonstrate that the proposed IMM-PPO framework effectively unifies predictive filtering, uncertainty-aware fusion, and reinforcement learning, enabling robust and intelligent automation for real-world logistics scenarios.