Reinforcement learning is one of the most natural formulations of the problem of acquiring knowledge in the interaction of an agent with the environment, however, there are many problems associated with efficiency and application in robotics. Usually, modern planners are assigned sets of actions in advance at the symbolic level (in the form of special rules), but for a real robot that may find itself in a completely new situation, this approach does not work well. He must recall his past experience in similar situations and draw up a new plan from new actions that no one prompted him.

In this direction, we develop both new methods and algorithms and solve new applied problems, including those with robotic manipulators. We also actively participate in competitions. For example, at the end of 2019. our team became the winner of the international NeurIPS MineRL competition, showing the best solution towards effective demonstration-based reinforcement learning methods.

Pages of completed projects in this direction:

• ForgER: reinforcement learning with forgetful demonstrations - https://sites.google.com/view/forgetful-experience-replay/home
• Reinforcement Learning to Explore the Environment in a Navigation Problem - https://cog-isa.github.io/rl_navigation/
• DSN: discrete model reinforcement learning - https://cog-isa.github.io/schema-rl/
• Reinforcement learning for a self-driving car at an intersection- https://martinsonmichael.github.io/CarRacing_agents/