Title: Enhancing Pose Control in Double-Ackermann Robots via Deep Reinforcement Learning Amidst Actuation Uncertainties

Abstract:

Deploying deep reinforcement learning (DRL) policies on physical hardware continues to pose significant challenges, primarily due to the gap between simulated and real-world dynamics. This study tackles these difficulties within the specific domain of double-Ackermann-steering mobile robots, whose non-holonomic constraints introduce complex maneuvering requirements. Leveraging the existing ManeuverNet DRL framework, we expand its scope from simple position control to comprehensive pose control, thereby increasing the complexity of the task. We also examine how uncertainties related to actuation affect the transferability of these policies.

Our findings indicate that relying on simplified actuation models during the training of the extended policy can severely hinder generalization. Specifically, under rigorous evaluation conditions, the success rate plummeted from 100% in the PyBullet simulator to just 25% in Gazebo. To overcome this performance bottleneck, we implemented a sim-to-sim-to-real methodology. This strategy involves integrating the actuation dynamics observed in Gazebo back into the PyBullet training environment. By employing multi-environment DRL techniques with the Soft Actor-Critic (SAC) and CrossQ algorithms, we developed policies that demonstrate robustness against modeling inaccuracies. This method effectively narrows the performance disparity between simulators, yielding a success rate of up to 92% in Gazebo and sustaining a 69% rate under stricter criteria. Crucially, these policies were successfully transferred to a physical robot without the need for further tuning.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC