Sim2Real-AD: Modular framework for transferring VLM-guided RL to real driving

Transferring artificial intelligence models trained in simulated environments to the real world remains one of the most complex challenges in robotics and autonomous driving. Traditionally, reinforcement learning (RL) policies guided by vision-language models (VLMs) have shown great potential in simulators like CARLA, but they fail when faced with the unpredictable conditions of a real vehicle. The conventional approach of closing the simulation-to-reality (sim-to-real) gap through manual adjustments is often fragile and costly. In this context, the Sim2Real-AD framework proposes a modular decomposition that separates the problem into two orthogonal axes: perception and dynamics mismatch, and task and geometry mismatch. This separation allows applying techniques such as a geometric observation bridge and a physics-aware action mapping, enabling a policy trained exclusively in CARLA to be transferred without real data to a full-size Ford E-Transit van. The result is autonomous driving in scenarios such as vehicle following, obstacle avoidance, and stop signs, all without the need for retraining in the physical world. This breakthrough opens the door for companies specialized in AI for business to integrate custom software solutions that combine realistic simulations with safe deployments. At Q2BSTUDIO we work on developing custom applications that leverage artificial intelligence and AI agents to automate critical processes, while our AWS and Azure cloud services ensure the scalability of these workloads. Additionally, monitoring these systems can be enhanced with business intelligence services like Power BI, offering real-time visibility into autonomous vehicle performance. Cybersecurity also plays a fundamental role in protecting communications between the simulator and the real vehicle. Ultimately, the Sim2Real-AD framework not only represents an academic milestone but also a practical roadmap for enterprise software companies to implement more efficient and sustainable intelligent driving solutions.