This paper proposes a novel navigation approach for Unmanned Tracked Vehicles (UTVs) using prior-based ensemble deep reinforcement learning, which fuses the policy of the ensemble Deep Reinforcement Learning (DRL) and Dynamic Window Approach (DWA) to enhance both exploration efficiency and deployment safety in unstructured off-road environments. First, by integrating kinematic analysis, we introduce a novel state and an action space that account for rugged terrain features and track–ground interactions. Local elevation information and vehicle pose changes over consecutive time steps are used as inputs to the DRL model, enabling the UTVs to implicitly learn policies for safe navigation in complex terrains while minimizing the impact of slipping disturbances. Then, we introduce an ensemble Soft Actor–Critic (SAC) learning framework, which introduces the DWA as a behavioral prior, referred to as the SAC-based Hybrid Policy (SAC-HP). Ensemble SAC uses multiple policy networks to effectively reduce the variance of DRL outputs. We combine the DRL actions with the DWA method by reconstructing the hybrid Gaussian distribution of both. Experimental results indicate that the proposed SAC-HP converges faster than traditional SAC models, which enables efficient large-scale navigation tasks. Additionally, a penalty term in the reward function about energy optimization is proposed to reduce velocity oscillations, ensuring fast convergence and smooth robot movement. Scenarios with obstacles and rugged terrain have been considered to prove the SAC-HP’s efficiency, robustness, and smoothness when compared with the state of the art.
Predicting a quadruped’s response to planned commands is difficult when terrain geometry and contact conditions differ from training. A terrain category associated with high traction during training may provide low traction at deployment, weakening predictions based on that association. We address this problem with the Calibration-Enhanced State-Transition Prediction Model (CE-STPM). A terrain branch aggregates category prototypes using training priors, while a motion branch combines a global interaction dictionary with features from recent motion history. These branches initialize a shared recurrent predictor of future positions and failure risk under candidate commands. Back-door and front-door reasoning motivate the reference design without identifying causal effects. In Unitree Go2 simulations in Isaac Lab, CE-STPM reduced mean final displacement error by 44.3% under a geometry shift relative to the backbone, STPM-Base. The reduction was 47.9% under a shift in contact conditions, where F1 for risk prediction increased from 0.75 to 0.86. Comparisons using individual branches supported complementary roles, and a controlled association test showed reduced degradation. Across 150 physical robot trajectory snippets with offline terrain priors, CE-STPM achieved lower displacement error and higher risk precision than the backbone in all three scene groups. These results support combining terrain and motion references before recurrent prediction.