Research

Research

My research interests include, but are not limited to:

  • Artificial Intelligence;
  • intelligent robotics;
  • AI planning
  • path/motion planning;
  • heuristic search;
  • multi-agent systems;
  • vision-based navigation;
  • cognitive agents.

I’m passionate and enthusiastic about developing methods and algorithms for controlling intelligent agents (mobile robots, driverless cars, drones, computer game characters etc.) in such a way that these agents are:

  • autonomous, i.e. can behave appropriately in complex, dynamic environments without being fully controlled by an operator;
  • adaptive, i.e. can behave well in changing environments;
  • collaborative, i.e. can interact with each other and humans to safely and effectively accomplish their missions.

Creating intelligent control systems for mobile agents is a challenging problem. To solve it one needs to use the variety of methods from AI, control theory, computer cognitive modelling etc.

Currently I’m involved in research and development of methods and algorithms for path and motion planning and (to a less extent) vision-based navigation.

Videos || Results || In plain English

Videos

The following videos provide an insight of my research activities.

Show description
This video demonstrates how our decentralized multi-agent planner based on reinforcement learning and heuristic search works. Each agent is moving towards its goal, and upon reaching it new one is released. This setting is called lifelong and it commonly appears in practice (e.g. in automated warehouses). Agents rely only on local observations while no communication is present. Despite such limitations agents are able to effectively navigate in the environment thanks to our method that integrates planning and learning.

Show description
On this video a mobile wheeled robot autonomously navigates to a goal set by an operator, while moving in uneven terrain with slopes (which is non-trivial task for wheeled robots). All computation is performed on-board in real time. The robot’s control system is composed of the three modules: 2.5D mapping, path planning and path following. All of these are implemented as ROS (Robotic Operating System) modules. Path planning and following are original and designed by us.

Show description
This video demonstrates the application of the single-agent path finding algorithm that takes into account both the kinodynamic constraints of the agent and the dynamic obstacles that populate the environment (their trajectories are considered to be accurately predicted and passed to the planner as the input).

Show description
This video demonstrates the application of the multi-agent path finding (MAPF) algorithm developed by us to various navigation scenarios in simulated and real-world environments. Unlike many other MAPF algorithms (present at that time) our approach allows actions of arbitrary duration and translations in arbitrary direction. Moreover it explicitly takes into account rotations as well. Being a prioritized planner it scales well to large number of robots (up to hundreds in simulation).

Another video on this line of research is here: https://www.youtube.com/watch?v=1Jrye5S0ZV8.

Show description
This video demonstrates how a map is built by processing the video-stream from a single camera (so-called monocular vSLAM). All processing is done in real time on the embedded computer NVidia Jetson. Fully-convolutional neural network of the original architecture is used to infer depth-maps which are further used by a conventional RGB-D SLAM method, i.e. RTAB-MAP. Resultant map is available both as a point-cloud (top-right of the frame) and as a voxel grid (bottom right) suitable for path planning.

Another video on this line of research is here: https://www.youtube.com/watch?v=piuVq8f61gs.
This video shows how the feature-based vSLAM method is used (ORB2-SLAM) for localization and mapping. Note how sparse the resultant map (red and black dots) is.

Show description
Demonstration of the autonomous flight in GPS-denied environment guided by computer vision. Drone detects markers on the floor (QR-codes) and follows instructions encoded with them, e.g. “pursue in the same direction”. When the landing marker is encountered drone automatically lands on it.

Results

Below one might find several examples of the algorithms, methods and models that were developed with my active involvement. More details on them can be found in my publications.

  1. Methods and algorithms for centralized multi-agent path finding (MAPF).
    • CCBS – an optimal MAPF planner that supports continuous time. A long (journal) version of this work is available here. (This is one of my most cited works)
    • ICCBS – an improved version of CCBS that significantly increases its computational efficiency without sacrificing theoretical guarantees.
    • AA-SIPP(m) – a prioritized multi-agent path planning algorithm that is capable of handling any-angle moves.
    • Enhanced AA-SIPP(m) algorithm that supports agents of arbitrary size, supports different moving velocities and takes into account rotations.
  2. Methods and algorithms for decentralized multi-agent path finding
    • Follower – method based on the combination of heuristic search (for long-term planning) and reinforcement learning (for local collision avoidance).
    • Switcher– one more variant of combining search and reinforcement learning for solving decentralized multi-agent pathfinding, that is based on different switching mechanisms that alternate between learnable and non-learnable behavior policies.
    • MATS-LP – decentralized multi-agent pathfinding based on the Monte-Carlo Tree Search (MCTS). One more paper on this topic (with the preliminary results).
  3. Methods and algorithms for single-agent path finding that utilize state-of-the-art machine learning techniques
    • POLAMP – algorithm for planning a path in a dynamic environment, which is based on the combination of reinforcement learning with search-based and sampling-based planning.
    • TransPath – algorithm for grid-based pathfinding that leverages a modern neural network (transformer) tailored and trained to (significantly) reduce the search space. A preliminary paper on this topic is available here.
    • GridPathRL – one of our first works on trying to solve classical pathfinding problem with reinforcement learning. (This is one of my most cited works)
  4. Methods and algorithms for single-agent path finding that are based on heuristic search
    • SIPP-IP – algorithm for pathfinding in dynamic environments that is based on the safe interval path planning methodology (for the sake of computational efficiency) and takes the kinodynamic constraints into account (e.g. the agent can not stop instantaneously).
    • TO-AA-SIPP – algorithm for pathfinding in dynamic environments that supports any-angle moves and guarantees the optimality of the constructed solutions.
    • Algorithm for finding a path that does not contain sharp turns – LIAN and its modifications: eLIAN and LP-LIAN.
  5. Methods and algorithms for vision-based navigation

In plain English

My peers and I are aimed at making mobile agents, e.g. mobile robots, smarter. Creating a intelligent, versatile robot is indeed a very challenging problem that is composed of dozens of smaller problems itself. We are focused on the software part of this puzzle. We create methods and algorithms, implement them as a software and evaluate them either in simulation (numerous our experiments look pretty much like “jumping dots on the screen”) or on the off-the-shelf robotic platforms (wheeled robots, drones, robotic arms etc.). We do not construct robots themselves, as we are more the software guys rather than the hardware guys. Moreover, some of our methods are general enough to be applied not only to physical robots but to, say, video game characters. Generally, as most of the ‘science guys’ we solve challenging puzzles that advance our understanding on how intelligent decision making should be implemented within the artificial systems.