Uploaded April 2024 | Updated September 2026, 2 hours ago
Microsoft Future Leaders in Robotics and AI Seminar Series: Robot Safety and Generalization in the Era of Foundation Models
Online Seminar
Anushri Dixit
Postdoctoral Fellow
Princeton University
Significant strides in AI over the past few years have enabled robotic systems to interpret and interact with the world in increasingly versatile ways. The large, often multi-modal, datasets that are used to train modern learning-based systems endow robots with capabilities like scene understanding and commonsense reasoning. However, the safe integration and reliability of these learned models for robotics applications still remains in question. Learned perception systems fail to identify objects correctly and LLM-based planners hallucinate their outputs leading to unsafe robot behavior downstream. In this talk, I will discuss a technique called conformal prediction and its usefulness for quantifying the uncertainty of such learned models. First, I will discuss a framework for rigorously quantifying the uncertainty of a pre-trained obstacle detection system in a way that provides a formal assurance on correctness and safety for planning applications. Next, I will present a LLM-based planning framework wherein given a human instruction, the robot is statistically guaranteed to complete the task while asking for human help if it is uncertain. I will provide the experimental validation of these methods on various robotic platforms for navigation and mobile manipulation tasks.
For more information on the Maryland Robotics Center see:
https://robotics.umd.edu
Microsoft Future Leaders in Robotics and AI Seminar Series: Robot Safety and Generalization in the Era of Foundation Models
Online Seminar
Anushri Dixit
Postdoctoral Fellow
Princeton University
Significant strides in AI over the past few years have enabled robotic systems to interpret and interact with the world in increasingly versatile ways. The large, often multi-modal, datasets that are used to train modern learning-based systems endow robots with capabilities like scene understanding and commonsense reasoning. However, the safe integration and reliability of these learned models for robotics applications still remains in question. Learned perception systems fail to identify objects correctly and LLM-based planners hallucinate their outputs leading to unsafe robot behavior downstream. In this talk, I will discuss a technique called conformal prediction and its usefulness for quantifying the uncertainty of such learned models. First, I will discuss a framework for rigorously quantifying the uncertainty of a pre-trained obstacle detection system in a way that provides a formal assurance on correctness and safety for planning applications. Next, I will present a LLM-based planning framework wherein given a human instruction, the robot is statistically guaranteed to complete the task while asking for human help if it is uncertain. I will provide the experimental validation of these methods on various robotic platforms for navigation and mobile manipulation tasks.
For more information on the Maryland Robotics Center see:
https://robotics.umd.edu
![Maryland Robotics Center Seminar, January 31, 2025: Markus P. Nemitz [REUPLOAD]](https://i.ytimg.com/vi/g1aHjIuysYI/mqdefault.jpg "Maryland Robotics Center Seminar, January 31, 2025: Markus P. Nemitz [REUPLOAD]
Maryland Robotics Center Seminar: Design. Print. Deploy. Enabling Swarm Robotics via Additive Manufacturing
Markus P. Nemitz, Ph.D.
Assistant Professor, Mechanical Engineering
Tufts University
Swarm robots offer transformative potential for applications where rapid and efficient coverage of large areas is critical. However, the high cost and fragility of advanced robots, coupled with the limited functionality of affordable alternatives, have historically hindered their large-scale deployment, confining much of swarm robotics research to simulations. While drones have successfully evolved into capable, low-cost swarm robots through commercialization, their fragility and inability to physically interact with the environment have limited their use in contact-based tasks and underwater operations. To address these challenges, our research focuses on the rapid design and fabrication of low-cost, capable, and scalable terrestrial swarm robots using additive manufacturing. In this talk, I will present strategies for creating increasingly intelligent yet affordable robots, advancing the cost-capability trade-off in robotics. I will showcase developments in 3D-printed soft quadrupeds, fluidic actuators, and controllers, exploring their current performance, future potential, and how they bridge the gap between laboratory prototypes and real-world applications. By advancing swarm engineering, we unlock new opportunities for distributed problem solving, such as explosive ordnance disposal, where expendable robots can be produced at a cost lower than the landmines they neutralize.
For more information on the Maryland Robotics Center see:
https://robotics.umd.edu")
