Embedded Files
The seminar series is broadly focused on the intersection of control and machine learning. It covers a wide range of topics including, but not limited to, reinforcement learning, learning for dynamical systems, optimal control and model-predictive control, online learning and adaptive control, interactive decision-making, control-theoretic perspectives on deep learning, and applications in robotics and other various real-world systems.
Next upcoming seminar
Next upcoming seminar
TBD
Past Seminar
Past Seminar
Ravi Pandya
Ravi Pandya
RI PhD at CMU
May 1, 2024
Towards Influence-Aware Safe Human-Robot Interaction
May 1, 2024
Towards Influence-Aware Safe Human-Robot Interaction
Video: https://youtu.be/-NKABCGJ-cQ
Bio: Ravi Pandya is a 4th year PhD student in the Robotics Institute at Carnegie Mellon University advised by Changliu Liu and Andrea Bajcsy. His work focuses on the intersection of safe control, human-robot interaction, and learning for robotics in order to make safe, adaptive, collaborative algorithms for robots to work with people. He holds a B.S. in EECS from the University of California, Berkeley and is a recipient of the NSF Graduate Research Fellowship.
Yufei Wang
Yufei Wang
RI PhD at CMU
April 17, 2024
Differentiable trajectory optimization as a policy representation for reinforcement and imitation learning
April 17, 2024
Differentiable trajectory optimization as a policy representation for reinforcement and imitation learning
Video: https://youtu.be/6wc_YnOemxQ
Bio: Yufei Wang is a third year Phd student at Robotics Institute, Carnegie Mellon University, co-advised by Prof. Zackory Erickson and Prof. David Held.
He received M.S. in Computer Science from Computer Science Department, Carnegie Mellon University in Dec, 2020, advised by Prof. David Held, and B.S. in Data Science from Yuanpei College, Peking University in July 2019, advised by Prof. Bin Dong. His general research interest is robot learning. His graduate study is supported by the Uber Presidential Fellowship.
Arundhati Banerjee
Arundhati Banerjee
ML PhD at CMU
April 3, 2024
Uncertainty-aware Decision Making in Multi-Agent Active Searc
April 3, 2024
Uncertainty-aware Decision Making in Multi-Agent Active Searc
Bio: Arundhati Banerjee is a PhD student in the Machine Learning department at CMU, advised by Prof. Jeff Schneider. During her PhD, Arundhati has worked on adaptive decision making algorithms in decentralized and asynchronous multi-agent systems for robotics search and tracking applications under realistic sensing, communication and resource considerations. Her research interests include active learning, Bayesian decision making, generative modeling for planning and multi-agent reinforcement learning.
March 13, 2024
Real-Time Safe Control of Neural Network Dynamic Models with Sound Approximation
March 13, 2024
Real-Time Safe Control of Neural Network Dynamic Models with Sound Approximation
Video: https://www.youtube.com/watch?v=zdbhPZ0Epow&ab_channel=LeCARLabatCMU
Abstract: Safe control of neural network dynamic models (NNDMs) is important to robotics and many applications. However, it remains challenging to compute an optimal safe control in real time for NNDM. In this talk, I will talk about our recent work to use a sound approximation of the NNDM in the control synthesis. In particular, we propose Bernstein over-approximated neural dynamics (BOND) based on the Bernstein polynomial over-approximation (BPO) of ReLU activation functions in NNDM. To mitigate the errors introduced by the approximation and to ensure the persistent feasibility of the safe control problems, we synthesize a worst-case safety index using the most unsafe approximated state within the BPO relaxation of NNDM offline. For the online real-time optimization, we formulate the first-order Taylor approximation of the nonlinear worst-case safety constraint as an additional linear layer of NNDM with the ℓ2 bounded bias term for the higher-order remainder. Comprehensive experiments with different neural dynamics and safety constraints show that with safety guaranteed, our NNDMs with sound approximation are 10-100 times faster than the safe control baseline that uses mixed integer programming (MIP), validating the effectiveness of the worst-case safety index and scalability of the proposed BOND in real-time large-scale settings.
Bio: Hanjiang Hu is a PhD student in the Department of Electrical and Computer Engineering (ECE) and Master of Machine Learning (MSML) student at Carnegie Mellon University, advised by Prof. Changliu Liu at the Robotics Institute. His research focuses on safety and robustness at the intersection of robotics, control theory, and machine learning, from the robustness certification of robot perception to the formal verification of neural networks for safe autonomous systems and robotics with provable guarantees. He obtained his bachelor's and master's degrees at Shanghai Jiao Tong University.
Feb 21, 2024
Feb 21, 2024
Efficient Reductions for Interactive Learning from Implicit Human Feedback
Efficient Reductions for Interactive Learning from Implicit Human Feedback
Video: https://www.youtube.com/watch?v=zdbhPZ0Epow&ab_channel=LeCARLabatCMU
Slides: https://drive.google.com/file/d/1So1ZorDB3y0dSLb09iAVOO2PWYvV897W/view?usp=sharing
Abstract: Interactive approaches to imitation learning (i.e. inverse reinforcement learning -- IRL) and learning from human preferences (i.e. RL from Human Feedback -- RLHF) have become the preferred approaches for challenging problems that range from autonomous driving to large language model fine-tuning. Despite their impressive empirical performance and strong theoretical guarantees, interactive approaches to learning from implicit feedback often come with a high computational burden. In this talk, we will discuss how adopting a game-theoretic perspective on these problems can allow us to derive efficient reductions that maintain the benefits of such approaches while dramatically reducing the amount of computation required, both in theory and practice. Specifically, we will talk about a new paradigm for IRL that does not require repeatedly solving a hard exploration / RL problem in the inner loop and a new flavor of algorithm for RLHF that avoids reward modeling and adversarial training while providing robustness to the intransitive preferences that often appear when aggregating diverse human judgements.
Bio: Gokul Swamy is a 4th year PhD student in the Robotics Institute at Carnegie Mellon University, where he works with Drew Bagnell and Steven Wu. His research centers on efficient algorithms for interactive learning from implicit human feedback (e.g. imitation learning, reinforcement learning from human feedback, learning with limited state information) and builds on techniques from RL, game theory, and causal inference. He has spent summers at Google Research, MSR, NVIDIA, Aurora, and SpaceX and holds M.S. / B.S. degrees from UC Berkeley.
Page updated
Report abuse