The spring series of seminars organized by our Center will include two directions: a review of current work from leading conferences on the topics of our laboratories and presentations by students and graduate students with the results (final and intermediate) of the projects in which they participate. On the topic of the seminars, we conditionally distinguish four topics: reinforcement learning (RL), computer vision (CV), behavior planning and control (Plan, Control), general artificial intelligence (Sign, Neuro).
Our seminars are open - we invite everyone to take part. If you have a topic that fits into our directions, write to our contacts - we are glad to expand the community and will include all worthy reports in the program.
Our seminars are open - we invite everyone to take part. If you have a topic that fits into our directions, write to our contacts - we are glad to expand the community and will include all worthy reports in the program.
График
Planning for a Robotic Platform with Kinematic Constraints | Brian Angulo, Zane Ali, Vladislav Golovin
Thu, 11 March 17:00
The report will consider various methods for solving one of the fundamental problems of mobile robotics - the problem of planning the trajectory of a robot's motion, taking into account kinematic constraints. Taking these constraints into account significantly complicates the task and makes its solution much less trivial than, for example, geometric planning. Nevertheless, as we will show in the report, there are techniques that can solve the problem in this setting within a reasonable time.
The report will consider various methods for solving one of the fundamental problems of mobile robotics - the problem of planning the trajectory of a robot's motion, taking into account kinematic constraints. Taking these constraints into account significantly complicates the task and makes its solution much less trivial than, for example, geometric planning. Nevertheless, as we will show in the report, there are techniques that can solve the problem in this setting within a reasonable time.
| Запись |
Analysis of approaches to the generation of high-dimensional semantic maps based on on-board sensors of unmanned vehicles | Alexey Karmanov
Thu, 4 March 17:00
The seminar will cover the purpose of HD maps, the main methods of their creation, as well as the generation of HD maps using onboard cameras. In the second part, a significant portion of attention will be paid to online mapping and recognition of dynamic objects (the so-called Bird-eye view segmentation task)
The seminar will cover the purpose of HD maps, the main methods of their creation, as well as the generation of HD maps using onboard cameras. In the second part, a significant portion of attention will be paid to online mapping and recognition of dynamic objects (the so-called Bird-eye view segmentation task)
| Запись |
Overview of Reinforcement Learning works
conference ICLR 2021 | Andrey Gorodetsky
conference ICLR 2021 | Andrey Gorodetsky
Thu, 25 February 17:00
The International Conference on Learning Representations (ICLR) is a conference dedicated to the development of a field of artificial intelligence called representative learning, but commonly referred to as deep learning.
Along with ICML and NeurIPS, ICLR is one of the three major conferences on Machine Learning and Artificial Intelligence and is continually recruiting a large number of articles. This year the conference accepted 860 out of 2997 papers.ICLR 2021 will take place from May 3-7, building on last year's experience, will be fully translated online due to the pandemic.
The seminar will review the work of the conference with an emphasis on unsupervised and meta reinforcement learning. Using these approaches, you can use the experience of the agent's past interaction with the environment or a set of suboptimal trajectories of the environment to accelerate the agent's tuning for new tasks. Also, in the absence of a reward function, it is possible to stabilize dynamical systems or acquire potentially useful skills for solving future problems / teaching the agent's behavior.
The International Conference on Learning Representations (ICLR) is a conference dedicated to the development of a field of artificial intelligence called representative learning, but commonly referred to as deep learning.
Along with ICML and NeurIPS, ICLR is one of the three major conferences on Machine Learning and Artificial Intelligence and is continually recruiting a large number of articles. This year the conference accepted 860 out of 2997 papers.ICLR 2021 will take place from May 3-7, building on last year's experience, will be fully translated online due to the pandemic.
The seminar will review the work of the conference with an emphasis on unsupervised and meta reinforcement learning. Using these approaches, you can use the experience of the agent's past interaction with the environment or a set of suboptimal trajectories of the environment to accelerate the agent's tuning for new tasks. Also, in the absence of a reward function, it is possible to stabilize dynamical systems or acquire potentially useful skills for solving future problems / teaching the agent's behavior.
| Запись |
MPC in manipulator control | Konstantin Mironov
Thu, 18 February 17:00
Model Predictive Control is currently a common method for adaptive control of robotic arms. A significant difficulty in applying this method is that linear process models are used in MPC, while standard models of manipulator dynamics are nonlinear and, accordingly, require linearization. The report examines modern approaches to the use of MPC to control robotic manipulators.
Model Predictive Control is currently a common method for adaptive control of robotic arms. A significant difficulty in applying this method is that linear process models are used in MPC, while standard models of manipulator dynamics are nonlinear and, accordingly, require linearization. The report examines modern approaches to the use of MPC to control robotic manipulators.
| Запись |
Reinforcement Learning Work Review NeurIPS 2020 v2.0
Thu, 11 February 17:00
We continue to get acquainted with the work on reinforcement learning from the NeurIPS 2020 conference. Last time Artem focused on articles devoted to model-based RL, today the guys will talk about articles from other areas of reinforcement learning.
We continue to get acquainted with the work on reinforcement learning from the NeurIPS 2020 conference. Last time Artem focused on articles devoted to model-based RL, today the guys will talk about articles from other areas of reinforcement learning.
| Запись |
Features of the implementation of CV and SLAM algorithms in ROS2 | Yushaa Murkhizh, Linar Abdrazakov, Ilya Belkin
Thu, 4 February 17:00
ROS (Robot Operating System) provides developers with libraries and tools to create robotics applications. ROS provides hardware abstraction, offers device drivers, libraries, renderers, messaging, package managers, and more. ROS is released under the terms of the BSD license and is open source.
A lot has changed in the robotics and ROS community since the launch of ROS in 2007. The goal of the ROS 2 project is to adapt to these changes by taking advantage of ROS 1 and improving what is not.
The seminar will demonstrate the features of ROS2 and its differences and advantages over ROS1 in solving computer vision and SLAM problems.
ROS (Robot Operating System) provides developers with libraries and tools to create robotics applications. ROS provides hardware abstraction, offers device drivers, libraries, renderers, messaging, package managers, and more. ROS is released under the terms of the BSD license and is open source.
A lot has changed in the robotics and ROS community since the launch of ROS in 2007. The goal of the ROS 2 project is to adapt to these changes by taking advantage of ROS 1 and improving what is not.
The seminar will demonstrate the features of ROS2 and its differences and advantages over ROS1 in solving computer vision and SLAM problems.
| Запись |
NeurIPS 2020 Reinforcement Learning Review: Environment and Planning Models | Artem Zholus
Thu, 28 January 17:00
NeurIPS is the premier conference on machine learning and computational neuroscience, where thousands of leading scientists and researchers gather each December to exchange research in the fields of cognitive sciences, psychology, computer vision, statistical linguistics, and information theory.
At our second seminar, Artem Zholus, MIPT master's student, will review the most interesting articles, and also highlight general trends in the model-based Reinforcement Learning direction.
NeurIPS is the premier conference on machine learning and computational neuroscience, where thousands of leading scientists and researchers gather each December to exchange research in the fields of cognitive sciences, psychology, computer vision, statistical linguistics, and information theory.
At our second seminar, Artem Zholus, MIPT master's student, will review the most interesting articles, and also highlight general trends in the model-based Reinforcement Learning direction.
| Запись |
Monocular Visual SLAM Techniques | Andrey Bokovoy, Kirill Muravyov
Analysis of neural network approaches to solving the Place Recognition problem | Yaroslav Solomentsev
Analysis of neural network approaches to solving the Place Recognition problem | Yaroslav Solomentsev
Thu, 21 January 17:00
An analysis of the capabilities of visual SLAM methods based on data from only one camera of a mobile robot will be given.
Various benchmarks of global localization by images are presented, including those considered at ECCV2020, as well as state-of-the-art methods for solving such a problem using neural networks.
An analysis of the capabilities of visual SLAM methods based on data from only one camera of a mobile robot will be given.
Various benchmarks of global localization by images are presented, including those considered at ECCV2020, as well as state-of-the-art methods for solving such a problem using neural networks.
| Запись |