Series
IRIM Seminar Series

Permanent Link
https://hdl.handle.net/1853/70926
Series Type
Event Series
Description
Associated Organization(s)
Associated Organization(s)

Filters

1
147
147
Reset filters

Settings
Type: Moving Image × Has files: Yes ×

Publication Search Results

Now showing 1 - 10 of 147
No Thumbnail Available
Item

Learning Coordinated Performant Flight with 20 Neurons

2025-04-09 , Sukhatme, Gaurav

We have recently demonstrated the possibility of learning controllers that are zero-shot transferable to groups of real quadrotors via large-scale, multi-agent, end-to-end reinforcement learning. We train policies parameterized by neural networks that can control individual drones in a group in a fully decentralized manner. Our policies, trained in simulated environments with realistic quadrotor physics, demonstrate advanced flocking behaviors, perform aggressive maneuvers in tight formations while avoiding collisions with each other, break and re-establish formations to avoid collisions with moving obstacles, and efficiently coordinate in pursuit-evasion tasks. The model learned in simulation transfers to highly resource-constrained physical quadrotors. Motivated by these results and the observation that neural control of memory-constrained, agile robots requires small yet highly performant models, the talk will conclude with some thoughts on coaxing learned models onto devices with modest computational capabilities.

View
Thumbnail Image
Item

Autonomous Systems in the Intersection of Control, Learning, and Formal Methods

2025-02-19 , Topcu, Ufuk

Autonomous systems are emerging as a driving technology for countlessly many applications. Numerous disciplines tackle the challenges toward making these systems trustworthy, adaptable, user-friendly, and economical. On the other hand, the existing disciplinary boundaries delay and possibly even obstruct progress. He argues that the nonconventional problems that arise in designing and verifying autonomous systems require hybrid solutions at the intersection of control, learning, and formal methods (among other disciplines). He presents examples of such hybrid solutions in the context of learning in sequential decision-making processes. These results offer novel means for effectively integrating physics-based, contextual, or structural prior knowledge into data-driven learning algorithms. They improve data efficiency by several orders of magnitude and generalizability to environments and tasks the system had not previously experienced. He concludes with remarks on a few promising future research directions.

View
Thumbnail Image
Item

Robots that know when they don't know

2025-01-08 , Majumdar, Anirudha

Foundation models from machine learning have enabled rapid advances in perception, planning, and natural language understanding for robots. However, current systems lack any rigorous assurances when required to generalize to novel scenarios. For example, perception systems can fail to identify or localize unfamiliar objects, and large language model (LLM)-based planners can hallucinate outputs that lead to unsafe outcomes when executed by robots. How can we rigorously quantify the uncertainty of machine learning components such that robots know when they don’t know and can act accordingly? In this talk, I will present our group’s work on developing principled theoretical and algorithmic techniques for providing formal assurances on learning-enabled robots that act based on rich sensory inputs (e.g., vision) and natural language instructions. The key technical insight is to leverage and extend powerful methods from conformal prediction and generalization theory for rigorous uncertainty quantification in a way that complements and scales with the growing capabilities of foundation models. I will present experimental validation of our methods for providing strong statistical guarantees on LLM planners that ask for help when they are uncertain, and for vision-based navigation and manipulation.

View
Thumbnail Image
Item

A symbiotic philosophy for bio-inspired robotics

2024-03-27 , Moore, Talia

Humans have frequently looked to natural phenomena to inspire the design of art, structures, and mechanisms. However, there are as many different ways to learn from nature as there are words for this approach: bioinspiration, biomimicry, and biodesign to name a few. In this talk, I propose a taxonomy for categorizing distinct biodesign approaches and use examples from my own research to illustrate the methodology and benefits of each. In particular, I introduce the field of Animal-Robot Interactions and describe how bio-inspired approaches can be used to further biological inquiry while advancing robotics.

View
Thumbnail Image
Item

Resilient Autonomy for Extreme and Uncertain Environments

2025-03-26 , Scherer, Sebastian

Robustness in AI and robotics shows great promise if systems can transition from the lab to real-world environments, moving beyond the single-operator per robot paradigm. However, the unstructured nature of the real-world demands nuanced decision-making from robots. In this talk, Prof. Scherer will outline approaches, progress, and results in multi-modal sensing, nuanced perception inputs, navigation in difficult terrain, and extensions to multi-robot teams, as well as future research directions.

View
Thumbnail Image
Item

Enhancing human mobility with agile robotic prostheses and orthoses

2025-02-05 , Gregg, Robert

Even with the help of modern prosthetic and orthotic devices, individuals with lower-limb amputation, age-related motor deficits, or orthopedic disorders often struggle to navigate the home and community. Emerging powered prosthetic and orthotic devices could actively assist individuals to enable greater mobility, but these devices are typically designed to produce a small set of pre-defined motions. Although the field is beginning to embrace controllers that unify phases of the gait cycle, these devices still switch between distinct controllers for different tasks, e.g., uphill vs. downhill. This discrete control paradigm cannot continuously synchronize the robot’s motion to the variable activities of the human user. This talk will first present a new paradigm for controlling powered prosthetic legs over continuous variations of walking and stairs (i.e., different speeds and inclines), as well as continuous transitions between sitting and standing. These adaptable mid-level controllers facilitate a small activity space for intent classification, enabling amputee users to control activity transitions through intuitive, heuristic rules with over 99% accuracy. While these methods reproduce missing joint function, a different control philosophy is needed for exoskeletons that assist existing joint function. The last part of this talk will introduce an energetic control paradigm for backdrivable exoskeletons to reduce muscular effort by providing a faction of the human torque, without requiring explicit knowledge of the activity. This task-agnostic control method enabled a bilateral knee exoskeleton to mitigate the effects of quadriceps fatigue in able-bodied individuals during repetitive lifting-lowering and carrying over 5 terrains, thus reducing their risk for injuries due to fatigue-induced compensations. The talk will conclude with preliminary results from studies using hip and knee exoskeletons to enhance the mobility of elderly individuals in real-world scenarios.

View
Thumbnail Image
Item

Unifying Semantic and Physical Intelligence for Generalist Humanoid Robots

2024-09-04 , Shi, Guanya

Humanoid robots offer two unparalleled advantages in general-purpose embodied intelligence. First, humanoids are built as generalist robots that can potentially do all the tasks humans can do in complex environments. Second, the embodiment alignment between humans and humanoids allows for the seamless integration of human cognitive skills with versatile humanoid capabilities. To build generalist humanoids, there are three critical aspects of intelligence: (1) Semantic intelligence (how the robot understands the world and reasons); (2) Physical/Motion intelligence (locomotion and manipulation skills); and (3) Mechanical/Hardware intelligence (how the robot actuates and senses). In this talk, I will present some recent works (H2O, OmniH2O, WoCoCo, ABS) that aim to unify semantic and physical intelligence for humanoid robots. In particular, H2O and OmniH2O provide a universal and dexterous interface that enables diverse human control (e.g., VR, RGB) and autonomy (e.g., using imitation learning or VLMs) methods for humanoids, WoCoCo provides an efficient framework for loco-manipulation skill learning without motion priors, and ABS provides safety guarantees for agile vision-based locomotion control. Finally, I will briefly discuss how to combine learning-based control approaches and traditional model-based control approaches to get the best of two worlds.

View
Thumbnail Image
Item

Computational Symmetry and Learning for Robotics

2025-03-05 , Ghaffari, Maani

Forthcoming mobile robots require efficient generalizable algorithms to operate in challenging and unknown environments without human intervention while collaborating with humans. Today, despite the rapid progress in robotics and autonomy, no robot can deliver human-level performance in everyday tasks and missions such as search and rescue, exploration, and environmental monitoring and conservation. In this talk, I will put forward a vision for enabling efficiency and generalization requirements of real-world robotics via computational symmetry and learning. I will walk you through structures that arise from combining symmetry, geometry, and learning in various foundational problems in robotics and showcase their performance in experiments ranging from perception to control. In the end, I will share my thoughts on promising future directions and opportunities based on lessons learned on the field and campus.

View
Thumbnail Image
Item

Hardware Design and Control Algorithms for Agile and Versatile Legged Robots

2025-01-22 , Park, Hae-Won

Biological animals excel at navigating complex and challenging environments, leveraging their hardware to perform dynamic motions and athletic skills that overcome diverse obstacles. Despite recent advancements, robotic systems still lack comparable dynamic locomotion capabilities. In this talk, I will present our lab's efforts to bridge this gap by developing robust hardware and effective control algorithms that enable both agility and robustness in legged robots. I will begin by introducing our quadruped robot platforms: HOUND, designed for high-speed locomotion on complex terrains, and MARVEL, designed for agile and versatile climbing. HOUND incorporates custom electric actuators, while MARVEL uses magnetic feet to generate climbing force. I will then discuss the control algorithms that drive these robots, leveraging model predictive control and reinforcement learning techniques. Finally, I will present our latest learning-based locomotion control framework, capable of synthesizing and executing diverse dynamic motions across various terrains. This framework combines a low-level skill policy, pre-trained using a large offline dataset generated via trajectory optimization, with a reinforcement learning policy trained on diverse terrains. With this integrated approach, HOUND achieves speeds of up to 9.5 m/s, making it the fastest legged robot, while MARVEL can traverse ceilings and vertical walls at speeds of up to 0.5 m/s and 0.7 m/s, respectively.

View
Thumbnail Image
Item

Regenstein Center for Bionic Medicine: Intuitive Control of Bionic Limbs

2024-04-10 , Hargrove, Levi

Amputation is a leading cause of disability, and prosthetic devices are commonly accepted treatment options to restore functional capabilities. However, current prosthetic devices still cannot fully match the functionality of their natural counterparts. This talk focuses on the progress made in the development and control of bionic limbs for individuals with limb loss. The first portion of the talk provides an overview of the development, testing and commercialization of pattern recognition control systems for prosthetic arms, including their operation with advanced surgical techniques, such as targeted muscle reinnervation. A significant emphasis of this work has been on evaluation based on real user feedback, ensuring that the developed technologies meet the actual needs and preferences of end users. The second portion focuses on the application of these approaches (ie statistical pattern recognition and finite state-machines) to controlling powered leg prostheses. Finally, I will discuss our recent work in using deep-learning coupled with benchmark datasets (some collected at Georgia Tech) to remove the reliance of finite-state machines from our overall control approach.

View