Article
Robotics
Zhangjie Cao, Dorsa Sadigh
Summary: The study addresses the issue of sub-optimal demonstrations in imitation learning and provides a metric to measure the usefulness of a demonstration. Experimental results show improved learned policies by utilizing the proposed score.
IEEE ROBOTICS AND AUTOMATION LETTERS
(2021)
Article
Computer Science, Information Systems
Dayong Xu, Fei Zhu, Quan Liu, Peiyao Zhao
Summary: The ARAIL algorithm addresses the issue of incomplete demonstrations by introducing a ranker model, reshaping the reward function, and improving performance and robustness on various levels of incompleteness in demonstrations.
INFORMATION SCIENCES
(2021)
Article
Robotics
Jihong Zhu, Michael Gienger, Jens Kober
Summary: This paper presents a novel approach to enhance the learning of task-parameterized skills by augmenting the training dataset with synthetic data. By utilizing this method, it is possible to train robots to perform flexible tasks with only a few demonstrations.
IEEE ROBOTICS AND AUTOMATION LETTERS
(2022)
Article
Automation & Control Systems
Yingbai Hu, Mingyang Cui, Jianghua Duan, Wenjun Liu, Dianye Huang, Alois Knoll, Guang Chen
Summary: Motion generation by imitating allows a robot to generate new trajectories in a different environment. Research on dynamic movement primitives (DMP) has shown promising results, but there are still challenges in learning from multiple demonstrations and incorporating obstacle avoidance. Combining DMP with model predictive control (MPC) can help address these issues.
ROBOTICS AND AUTONOMOUS SYSTEMS
(2023)
Article
Computer Science, Interdisciplinary Applications
Y. Q. Wang, Y. D. Hu, S. El Zaatari, W. D. Li, Y. Zhou
Summary: The research proposed an optimized approach to improve Gaussian Mixture Model and Gaussian Mixture Regression for enabling collaborative robots to effectively perform complex manufacturing tasks. The method includes a Gaussian noise strategy, an optimization algorithm, and integration of an interpolation algorithm, demonstrating good performance in terms of computational efficiency, solution quality, and adaptability.
ROBOTICS AND COMPUTER-INTEGRATED MANUFACTURING
(2021)
Article
Engineering, Electrical & Electronic
Ziye Hu, Wei Li, Zhongxue Gan, Weikun Guo, Jiwei Zhu, James Zhiqing Wen, Decheng Zhou
Summary: This study proposes a model-agnostic meta-learning framework based on task-contrastive learning to teach robots what to do and what not to do through positive and negative demonstrations. Experimental results demonstrate the effectiveness of this approach in simulated benchmarks and real-world experiments.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY
(2022)
Article
Computer Science, Artificial Intelligence
Jorge Ramirez, Wen Yu, Adolfo Perrusquia
Summary: Reinforcement learning from expert demonstrations (RLED) is a promising approach that combines imitation learning with reinforcement learning to improve sample efficiency in high-dimensional spaces. It considers prior knowledge and online knowledge as two possible sources to guide the reinforcement learning process. The survey focuses on novel methods for model-free reinforcement learning guided through demonstrations, and discusses challenges, applications, and promising approaches for improvement.
ARTIFICIAL INTELLIGENCE REVIEW
(2022)
Article
Automation & Control Systems
Rodrigo Perez-Dattari, Bruno Brito, Oscar de Groot, Jens Kober, Javier Alonso-Mora
Summary: This paper introduces a motion planning framework that combines data-driven policy and local trajectory optimization to achieve safe and socially compliant autonomous driving. By using interactive imitation learning, a model predictive controller is trained and validated in realistic simulated urban scenarios. The approach significantly improves navigation performance and expands the operational domain of the MPC to more realistic autonomous driving scenarios.
ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE
(2022)
Article
Robotics
Anna Meszaros, Giovanni Franzese, Jens Kober
Summary: This study investigates how to learn the intricate task of continuous pick & place motion from human demonstrations and corrections. The framework allows non-expert users to modify the dynamics of their initial demonstration through teleoperated corrective feedback, enabling them to teach motions beyond their physical capabilities. The framework learns the desired movement dynamics based on current Cartesian position with Gaussian Processes, allowing for online interactive corrections and active disturbance rejection.
IEEE ROBOTICS AND AUTOMATION LETTERS
(2022)
Article
Robotics
Yifeng Zhu, Peter Stone, Yuke Zhu
Summary: This article presents a method for tackling real-world long-horizon robot manipulation tasks through skill discovery. The method learns a library of reusable skills from unsegmented demonstrations and uses these skills to synthesize prolonged robot behaviors. The study demonstrates that skills discovered from multi-task demonstrations significantly improve the success rate of tasks compared to skills discovered from individual tasks.
IEEE ROBOTICS AND AUTOMATION LETTERS
(2022)
Article
Computer Science, Artificial Intelligence
Guoyu Zuo, Qishen Zhao, Shuai Huang, Jiangeng Li, Daoxiong Gong
Summary: The paper proposes a novel imitation learning algorithm, MD2-GAIL, which effectively learns from imperfect demonstrations and utilizes confidence scores to reconstruct the objective function for policy matching. By demonstrating better performance than other methods in experiments, the effectiveness of the algorithm is proven.
Article
Engineering, Multidisciplinary
JiaNan Yang, ShengAo Lu, MingHao Han, YunPeng Li, YuTing Ma, ZeFeng Lin, HaoWei Li
Summary: This paper addresses the problem of mapless navigation for unmanned aerial vehicles in scenarios with limited sensor accuracy and computing capability. It proposes a novel learning-based algorithm called soft actor-critic from demonstrations (SACfD), which integrates reinforcement learning with imitation learning. The algorithm utilizes maximum entropy reinforcement learning framework to enhance exploration capability and leverages demonstration data to accelerate convergence rate and improve policy performance. Experimental results demonstrate that the proposed algorithm outperforms existing algorithms.
SCIENCE CHINA-TECHNOLOGICAL SCIENCES
(2023)
Article
Computer Science, Artificial Intelligence
Fuxian Huang, Naye Ji, Huajian Ni, Shijian Li, Xi Li
Summary: In reinforcement learning, exploration is crucial but often inefficient in large state-action spaces or sparse rewards. To address this, a novel adaptive cooperative exploration method is proposed to alleviate the issues caused by imperfect demonstrations and improve policy learning.
PATTERN RECOGNITION LETTERS
(2023)
Article
Automation & Control Systems
Shirine El Zaatari, Weidong Li, Zahid Usman
Summary: This paper proposes a novel algorithm to address the issue of orientation restrictions in task parameter modelling by introducing ring Gaussians, and improves Gaussian mixture regression to generate regression paths adaptable to complex environments. Experimental results show that the improved algorithm outperforms the conventional one in terms of smoothness, efficiency, and reachability.
ROBOTICS AND AUTONOMOUS SYSTEMS
(2021)
Article
Computer Science, Information Systems
Lin Zhang, Fei Zhu, Xinghong Ling, Quan Liu
Summary: PUIWIL is an algorithm that utilizes non-negative positive-unlabeled learning to improve the utilization and performance of imitation learning from imperfect demonstrations. It evaluates the quality of expert demonstrations using confidence scores and reweights the demonstrations accordingly. PUIWIL also reconstructs the GAIL framework to enhance the impact of high-quality demonstrations on learning.
INFORMATION SCIENCES
(2022)