Grigory Yaremenko

Pavel Osinenko, Grigory Yaremenko, Roman Zashchitin et al.

This work presents and showcases a novel reinforcement learning agent called Critic As Lyapunov Function (CALF) which is model-free and ensures online environment, in other words, dynamical system stabilization. Online means that in each learning episode, the said environment is stabilized. This, as demonstrated in a case study with a mobile robot simulator, greatly improves the overall learning performance. The base actor-critic scheme of CALF is analogous to SARSA. The latter did not show any success in reaching the target in our studies. However, a modified version thereof, called SARSA-m here, did succeed in some learning scenarios. Still, CALF greatly outperformed the said approach. CALF was also demonstrated to improve a nominal stabilizer provided to it. In summary, the presented agent may be considered a viable approach to fusing classical control with reinforcement learning. Its concurrent approaches are mostly either offline or model-based, like, for instance, those that fuse model-predictive control into the agent.

Grigory Yaremenko, Dmitrii Dobriborsci, Roman Zashchitin et al.

Reinforcement Learning (RL) has been shown to be effective and convenient for a number of tasks in robotics. However, it requires the exploration of a sufficiently large number of state-action pairs, many of which may be unsafe or unimportant. For instance, online model-free learning can be hazardous and inefficient in the absence of guarantees that a certain set of desired states will be reached during an episode. An increasingly common approach to address safety involves the addition of a shielding system that constrains the RL actions to a safe set of actions. In turn, a difficulty for such frameworks is how to effectively couple RL with the shielding system to make sure the exploration is not excessively restricted. This work presents a novel safe model-free RL agent called Critic As Lyapunov Function (CALF) and showcases how CALF can be used to improve upon control baselines in robotics in an efficient and convenient fashion while ensuring guarantees of stable goal reaching. The latter is a crucial part of safety, as seen generally. With CALF all state-action pairs remain explorable and yet reaching of desired goal states is formally guaranteed. Formal analysis is provided that shows the goal stabilization-ensuring properties of CALF and a set of real-world and numerical experiments with a non-holonomic wheeled mobile robot (WMR) TurtleBot3 Burger confirmed the superiority of CALF over such a well-established RL agent as proximal policy optimization (PPO), and a modified version of SARSA in a few-episode setting in terms of attained total cost.

Grigory Yaremenko, Georgiy Malaniya, Pavel Osinenko

Online reinforcement learning is concerned with training an agent on-the-fly via dynamic interaction with the environment. Here, due to the specifics of the application, it is not generally possible to perform long pre-training, as it is commonly done in off-line, model-free approaches, which are akin to dynamic programming. Such applications may be found more frequently in industry, rather than in pure digital fields, such as cloud services, video games, database management, etc., where reinforcement learning has been demonstrating success. Online reinforcement learning, in contrast, is more akin to classical control, which utilizes some model knowledge about the environment. Stability of the closed-loop (agent plus the environment) is a major challenge for such online approaches. In this paper, we tackle this problem by a special fusion of online reinforcement learning with elements of classical control, namely, based on the Lyapunov theory of stability. The idea is to start the agent at once, without pre-training, and learn approximately optimal policy under specially designed constraints, which guarantee stability. The resulting approach was tested in an extensive experimental study with a mobile robot. A nominal parking controller was used as a baseline. It was observed that the suggested agent could always successfully park the robot, while significantly improving the cost. While many approaches may be exploited for mobile robot control, we suggest that the experiments showed the promising potential of online reinforcement learning agents based on Lyapunov-like constraints. The presented methodology may be utilized in safety-critical, industrial applications where stability is necessary.

Anton Bolychev, Georgiy Malaniya, Grigory Yaremenko et al.

We introduce a universal policy wrapper for reinforcement learning agents that ensures formal goal-reaching guarantees. In contrast to standard reinforcement learning algorithms that excel in performance but lack rigorous safety assurances, our wrapper selectively switches between a high-performing base policy -- derived from any existing RL method -- and a fallback policy with known convergence properties. Base policy's value function supervises this switching process, determining when the fallback policy should override the base policy to ensure the system remains on a stable path. The analysis proves that our wrapper inherits the fallback policy's goal-reaching guarantees while preserving or improving upon the performance of the base policy. Notably, it operates without needing additional system knowledge or online constrained optimization, making it readily deployable across diverse reinforcement learning architectures and tasks.

Georgiy Malaniya, Anton Bolychev, Grigory Yaremenko et al.

We introduce Multi-CALF, an algorithm that intelligently combines reinforcement learning policies based on their relative value improvements. Our approach integrates a standard RL policy with a theoretically-backed alternative policy, inheriting formal stability guarantees while often achieving better performance than either policy individually. We prove that our combined policy converges to a specified goal set with known probability and provide precise bounds on maximum deviation and convergence time. Empirical validation on control tasks demonstrates enhanced performance while maintaining stability guarantees.

Pavel Osinenko, Grigory Yaremenko, Ilya Osokin

Reinforcement learning is a general methodology of adaptive optimal control that has attracted much attention in various fields ranging from video game industry to robot manipulators. Despite its remarkable performance demonstrations, plain reinforcement learning controllers do not guarantee stability which compromises their applicability in industry. To provide such guarantees, measures have to be taken. This gives rise to what could generally be called stabilizing reinforcement learning. Concrete approaches range from employment of human overseers to filter out unsafe actions to formally verified shields and fusion with classical stabilizing controllers. A line of attack that utilizes elements of adaptive control has become fairly popular in the recent years. In this note, we critically address such an approach in a fairly general actor-critic setup for nonlinear time-continuous environments. The actor network utilizes a so-called robustifying term that is supposed to compensate for the neural network errors. The corresponding stability analysis is based on the value function itself. We indicate a problem in such a stability analysis and provide a counterexample to the overall control scheme. Implications for such a line of attack in stabilizing reinforcement learning are discussed. Furthermore, unfortunately the said problem possess no fix without a substantial reconsideration of the whole approach. As a positive message, we derive a stochastic critic neural network weight convergence analysis provided that the environment was stabilized.

Pavel Osinenko, Dmitrii Dobriborsci, Grigory Yaremenko et al.

A common setting of reinforcement learning (RL) is a Markov decision process (MDP) in which the environment is a stochastic discrete-time dynamical system. Whereas MDPs are suitable in such applications as video-games or puzzles, physical systems are time-continuous. A general variant of RL is of digital format, where updates of the value (or cost) and policy are performed at discrete moments in time. The agent-environment loop then amounts to a sampled system, whereby sample-and-hold is a specific case. In this paper, we propose and benchmark two RL methods suitable for sampled systems. Specifically, we hybridize model-predictive control (MPC) with critics learning the optimal Q- and value (or cost-to-go) function. Optimality is analyzed and performance comparison is done in an experimental case study with a mobile robot.