ENVIA - Simulation environment for AI-driven vehicles

我们如何学习安全的概念，如何将其应用于人工智能驾驶汽车的控制算法？虽然深度强化学习是现代人工智能在自主控制（机器人、飞机、工厂等）方面的一种非常常见的方法，但它本质上带有一个基本的安全风险，阻碍了它在许多实际工业系统中的部署。一些研究人员正在试图验证人工智能驱动系统的安全性。在这个项目中，我们使用了一种不同的方法：我们使用经典的控制方法来约束基于人工智能的系统，并防止不安全的行为。我们将控制理论和强化学习相结合，产生高效安全的控制算法，可以在模拟中进行初始训练，然后通过在目标真实硬件上满足安全保证进行调整和完善。其原理是使用“屏障”为通过强化学习训练的控制器（例如无人机控制器）设置严格的安全限制。这些障碍本身可以自动学习，并根据需要放松或收紧，从而在一定程度上降低风险以提高学习或表现。想象一下，一个人和教练一起学习滑雪：他们会冒着摔倒的风险去学习，但这种风险是有限的，受到教练的知识和指导的限制。同样地，我们的屏障将防止人工智能采取过于冒险的行动。我们的最终目标是设计可以在模拟中训练的控制器，然后部署在真实硬件上进行适应和额外的学习，并根据性能要求进行可调，计算风险。从中期来看，工业合作伙伴的经济和技术效益将是显著的，因为该项目将允许深度强化学习方法的应用，目前超过所有其他已知方法的性能，以及在越来越多的虚拟环境中，错误不会造成严重后果的人类专家，到安全至关重要的真实系统。从事该项目的人员将受到高度追捧，加拿大将受益于科学和工业声望，以及新的高素质人才的经济影响。

How do we learn the concepts of safety and how do we apply them to control algorithms for AI-driven vehicles? While deep reinforcement learning is a very common approach for modern artificial intelligence in terms of autonomous control (of robots, aircraft, factories, etc.), it carries by nature a fundamental safety risk that prevents its deployment in the context of many real industrial systems. Several researchers are trying to validate the safety of AI-driven systems. In this project, we use a different approach: we use classical control approaches to constrain AI-based systems and prevent actions that are not safe. We combine control theory and reinforcement learning to produce efficient and safe control algorithms, which can be initially trained in simulation, and then adapted and refined by meeting safety guarantees on the targeted real hardware. The principle is to use "barriers" to set strict safety limits for controllers trained by reinforcement learning (e.g., the controller of a drone). These barriers can themselves be learned automatically, and loosened or tightened as needed, trading off some level of risk to improve learning or performance. Imagine a person learning to ski with an instructor: they will take the risk of falling to learn, but this risk is limited and constrained by the instructor's knowledge and guidance. In a similar way, our barriers will prevent excessively risky actions taken by the AI. Our final goal is to design controllers that can be trained in simulation and then deployed on real hardware for adaptation and additional learning with tunable, calculated risk based on performance requirements. The economic and technical benefits for the industrial partners will be significant in the medium term, as the project will allow the application of deep reinforcement learning methods, which currently exceed the performance of all other known methods as well as that of human experts in more and more virtual environments where errors have no serious consequences, to real systems where safety is critical. Personnel working on the project will be highly sought after, and Canada will benefit in terms of scientific and industrialprestige, as well as by the economic impact of the new highly qualified personnel.