BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Robotics Institute Carnegie Mellon University - ECPv6.15.12.1//NONSGML v1.0//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:Robotics Institute Carnegie Mellon University
X-ORIGINAL-URL:https://www.ri.cmu.edu
X-WR-CALDESC:Events for Robotics Institute Carnegie Mellon University
REFRESH-INTERVAL;VALUE=DURATION:PT1H
X-Robots-Tag:noindex
X-PUBLISHED-TTL:PT1H
BEGIN:VTIMEZONE
TZID:America/New_York
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:20250309T070000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:20251102T060000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:20260308T070000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:20261101T060000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:20270314T070000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:20271107T060000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20260601T143000
DTEND;TZID=America/New_York:20260601T153000
DTSTAMP:20260711T161020
CREATED:20260526T132047Z
LAST-MODIFIED:20260526T132047Z
UID:151385-1780324200-1780327800@www.ri.cmu.edu
SUMMARY:Hierarchical Sub-Goal Policies for Generalizing Robot Manipulation
DESCRIPTION:Abstract:\nImitation learning has emerged as a leading paradigm for teaching manipulation skills to robots\, but its success depends on the costly endeavour of collecting robot demonstrations through teleoperation. Generalizing to novel objects\, environments\, and task variations typically requires massive datasets that are expensive to scale. This thesis investigates an alternative lever: hierarchy—explicitly factorizing manipulation policies into high-level reasoning about where to go and low-level control about how to get there—and asks whether such a factorization can both improve data efficiency from robot demonstrations alone and enable the use of\ncheaper sources of data such as human video.\n\nWe present GHOST\, a framework for learning visuomotor manipulation policies that generalize beyond the training distribution. GHOST factorizes control into (i) a high-level policy that predicts the next sub-goal as a distribution over 3D end-effector poses from multi-view RGB-D observations\, and (ii) a low-level goal-conditioned controller that executes embodiment-specific actions. To condition image-based policies on 3D goals\, we introduce a simple spatial interface that projects predicted goals into the image plane and represents them as end-effector heatmaps. Across a suite of manipulation tasks\, this hierarchical factorization consistently improves performance and robustness compared to a flat Diffusion Policy. We further show that this hierarchical interface makes it easy to incorporate human demonstrations without relying on (noisy) action retargeting. Because sub-goals are largely embodiment-agnostic\, we train\nthe high-level policy on human video to specify how learned skills should be applied and composed\, while keeping the low-level policy trained purely on robot data. This hierarchy enables adaptation to novel objects and task variations using only a small number of human demonstrations.\n\nWe evaluate GHOST on a suite of manipulation tasks spanning pick-and-place\, cloth folding\, and tool-use\, demonstrating both improved in-distribution performance and meaningful cross-embodiment generalization from human video. We conduct ablations to further isolate the contributions of the hierarchical factorization\, and identify the visual domain gap between human and robot observations as a significant bottleneck for long-horizon generalization. \nCommittee:\nDavid Held (advisor)\nShubham Tulsiani\nAviral Kumar\nYilin Wu
URL:https://www.ri.cmu.edu/event/hierarchical-sub-goal-policies-for-generalizing-robot-manipulation/
LOCATION:Gates Hillman Center 4405
CATEGORIES:MSR Thesis Presentation,Student Talks
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20260604T110000
DTEND;TZID=America/New_York:20260604T123000
DTSTAMP:20260711T161020
CREATED:20260529T154504Z
LAST-MODIFIED:20260529T154504Z
UID:151398-1780570800-1780576200@www.ri.cmu.edu
SUMMARY:User Intent-Driven and Context-Aware Personalization for Assistive Exoskeletons
DESCRIPTION:Abstract:\nPersonalizing exoskeleton control to individual preferences is crucial for real world deployment. Data-driven approaches have enabled user-generalizable controllers\, yet conventional personalization methods optimize biomechanical cost functions over user preferences. Prior work shows that users can perceive and report their preferred parameters\, yet no lightweight method maps user intent to quantitative control parameter changes in real time. \nTo address this gap\, we present a vision-language model (VLM) guided human-exoskeleton interface that translates natural user feedback and egocentric visual context into parameter updates for a hip exoskeleton controller. Our framework uses an off-the-shelf VLM that performs high-level intent parsing\, while low-level action selection is offloaded to a contextual bandit. This bandit explores and exploits over two personalization parameters\, the magnitude of assistance and the offset timing\, while learning individual preferences over time. We use human feedback (Likert scale) ratings as rewards for the bandit\, iterating parameter updates until the user reports satisfaction (Likert ≥ 5). \nWe evaluated this framework across a five-task locomotion track\, spanning level ground\, ramps\, and stairs\, across two laps\, against a single-shot interpreter and a one-shot-feedback baseline. Our contextual bandit pipeline improved from default initializations within a single interaction\, raising satisfaction by an average of 3 Likert points in multi-rating loops. No-improvement personalization instances were eliminated by the second lap\, and user satisfaction interactions increased from 67% to 95% across laps. We demonstrate VLMs paired with contextual bandits as an effective framework for online\, context-aware\, user preference learning in real-time exoskeleton personalization. \n\nCommittee:\nDr. Inseung Kang (chair)\nDr. Hartmut Geyer\nMichelle Zhao
URL:https://www.ri.cmu.edu/event/user-intent-driven-and-context-aware-personalization-for-assistive-exoskeletons/
LOCATION:Newell-Simon Hall 3305
CATEGORIES:MSR Thesis Presentation,Student Talks
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20260611T123000
DTEND;TZID=America/New_York:20260611T133000
DTSTAMP:20260711T161020
CREATED:20260604T185218Z
LAST-MODIFIED:20260604T185218Z
UID:151483-1781181000-1781184600@www.ri.cmu.edu
SUMMARY:Red and Blue Teaming for Robust Manipulation under Geometric Variations
DESCRIPTION:Abstract:\nRobotic manipulation policies are typically evaluated on curated\, in-distribution test sets\, which offer limited insight into how these policies behave under plausible variation. One important source of this variation is geometric in nature\, arising from small changes in object geometry that quietly alter grasp affordances and contact dynamics. Rather than treating robustness as a property to be measured passively\, this thesis develops methods that use object geometry as a controllable axis along which a policy can be actively stress-tested and subsequently improved.\n\nWe first introduce Geometric Red-Teaming (GRT)\, a framework that automatically discovers CrashShapes: structurally valid\, user-constrained mesh deformations that induce catastrophic failures in pre-trained manipulation policies. GRT couples a Jacobian-field deformation model with a gradient-free\, simulator-in-the-loop optimization strategy\, allowing it to search over plausible object geometries while treating the policy being tested as a black box. Across insertion\, articulation\, and grasping tasks\, GRT consistently finds deformations that collapse policy performance\, exposing brittle failure modes that static benchmarks miss. We further show that fine-tuning on a small set of CrashShapes together with the nominal object\, a process we call blue-teaming\, improves task success by up to 60 percentage points on those shapes while preserving performance on the original object. We validate both red-teaming and blue-teaming on a real robot arm\, where simulated CrashShapes reduce task success from 90% to as low as 22.5%\, and blue-teaming recovers performance to as high as 90% on the corresponding real-world geometry.\n\nWhile GRT provides a direct evaluation tool for geometric robustness\, its simulator-in-the-loop search takes hours per object and therefore cannot supply CrashShapes for large-scale blue-teaming. To address this limitation\, we introduce CrashGen\, a point-cloud-conditioned diffusion model that amortizes this search by generating failure-inducing deformation controls in seconds rather than hours. A deterministic Jacobian-field operator then maps these controls to physically plausible deformations of the original mesh. Trained once on controls discovered by running GRT offline\, CrashGen generates failure-inducing variants for over 80% of held-out objects under simulation validation. Such variants can in turn serve as targeted training data: blue-teaming Contact-GraspNet on a mixture of original meshes and CrashGen variants improves grasp success on held-out objects by up to 21% in simulation and 30% on a real robot\, relative to a baseline fine-tuned only on the original meshes.\n\nTogether\, GRT and CrashGen show that object geometry can serve two complementary roles in making manipulation policies more reliable under geometric variation. GRT provides an active evaluation tool for identifying policy-critical geometric perturbations\, while CrashGen provides a constructive mechanism for turning those perturbations into scalable training data. Once an expensive diagnostic\, geometric failure discovery becomes a scalable source of targeted supervision for the very policies it stress-tests.\n\nCommittee:\nZackory Erickson (co-chair)\nDavid Held (co-chair)\nAndrea Bajcsy\nYufei Wang
URL:https://www.ri.cmu.edu/event/red-and-blue-teaming-for-robust-manipulation-under-geometric-variations/
LOCATION:NSH 4305
CATEGORIES:MSR Thesis Presentation,Student Talks
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20260618T100000
DTEND;TZID=America/New_York:20260618T110000
DTSTAMP:20260711T161020
CREATED:20260615T135404Z
LAST-MODIFIED:20260615T135404Z
UID:151596-1781776800-1781780400@www.ri.cmu.edu
SUMMARY:Towards Fine-Grained Diagnosis of GUI Agents
DESCRIPTION:Abstract: \nGraphical User Interface (GUI) agents need strong planning—what to do next—and grounding—where to click next—to solve user tasks. Yet these agents remain unreliable\, and standard metrics such as task success or next-action accuracy often obscure why they fail. In this talk\, I argue that reliable GUI agents require fine-grained diagnosis of core agentic capabilities\, studied through two complementary directions. \nFirst\, I examine planning failures in GUI agents. Existing work can often identify planning errors\, but rarely pinpoints which component failed. This is largely because there is no structured breakdown of the skills an agent must master to plan effectively. This thesis identifies several such skills and proposes synthetic data pipelines to diagnose failures across them. Surprisingly\, even the strongest agents often fail on basic skills\, such as temporal reasoning over webpage sequences. I further show that performance on these fine-grained planning skills correlates with downstream task success\, suggesting a cheaper way to predict agent performance without expensive evaluations. \nSecond\, I diagnose failures in GUI grounding. Traditional accuracy-based metrics do not reliably expose hidden failure modes\, and I argue that sensitivity-based metrics are needed to compare models. In particular\, I find that no model consistently clicks the same UI element\, such as a calendar button\, across diverse user scenarios\, apps\, and operating systems. This raises concerns about whether these models truly understand what they click. To operationalize this diagnosis\, I propose a novel data generation pipeline and a GUI grounding diagnosis agent that automatically generates diverse instructions\, uses model feedback\, and iteratively identifies failure cases across state-of-the-art grounding models. Results show a significant mismatch between model rankings under accuracy-based metrics and our sensitivity-based metrics\, highlighting the need for more reliable agent performance measures. \nTogether\, this thesis provides diagnostics that can guide model development and advance more robust\, trustworthy interactive AI systems. Our future work delineates how each of these diagnostics can serve as a simple intervention to improve existing GUI agent pipelines. \nCommittee: \nFernando De La Torre\, co-chair\nAndrea Bajcsy\, co-chair\nLouis-Philippe Morency\nYinong (Oliver) Wang
URL:https://www.ri.cmu.edu/event/towards-fine-grained-diagnosis-of-gui-agents/
LOCATION:3305 Newell-Simon Hall
CATEGORIES:MSR Thesis Presentation,Student Talks
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20260618T100000
DTEND;TZID=America/New_York:20260618T110000
DTSTAMP:20260711T161020
CREATED:20260615T135538Z
LAST-MODIFIED:20260615T135538Z
UID:151598-1781776800-1781780400@www.ri.cmu.edu
SUMMARY:Embodied Design\, Modeling\, and Optimization for Bio-Inspired Aquatic Robots
DESCRIPTION:Abstract:\nBio-inspired aquatic robots offer a promising route to agile and efficient locomotion in fluid environments\, where conventional rigid systems remain limited. In bio-inspired aquatic systems\, locomotion is not determined by actuation or control alone\, but instead emerges from tightly coupled interactions among body morphology\, distributed compliance\, actuation\, onboard sensing\, and the surrounding flow\, making analysis\, design\, and optimization fundamentally challenging. \nThis thesis studies these coupled problems through the development of embodied aquatic robot platforms\, variational dynamical models\, and integrated sensing strategies\, with particular emphasis on undulatory and flapping swimming\, variable-stiffness design\, strongly coupled fluid–robot interaction\, and flow sensing and feedback based on local pressure measurements. The goal is to understand how morphology\, dynamics\, and sensing can be jointly exploited within a continuous-time dynamical framework to realize robotic swimmers that are more efficient\, maneuverable\, and adaptive\, while also enabling reliable optimization of physical design\, sensing configuration\, and control through optimization-based fluid dynamic formulations and high-fidelity gradients. These ideas are further examined and validated through experiments on physical robotic platforms. \nCommittee:\nProf. Carmel Majidi\nProf. Zachary Manchester\nProf. Victoria Webster-Wood\nDr. Jeong Hun Lee
URL:https://www.ri.cmu.edu/event/embodied-design-modeling-and-optimization-for-bio-inspired-aquatic-robots/
LOCATION:Gates Hillman Center 4405
CATEGORIES:MSR Thesis Presentation,Student Talks
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20260622T110000
DTEND;TZID=America/New_York:20260622T120000
DTSTAMP:20260711T161020
CREATED:20260616T135241Z
LAST-MODIFIED:20260616T135241Z
UID:151605-1782126000-1782129600@www.ri.cmu.edu
SUMMARY:Towards Generalizable Embodied Navigation with Vision-Language Models
DESCRIPTION:Abstract:\nEmbodied navigation asks an autonomous agent to move through unknown environments and accomplish tasks such as finding objects or following instructions. Reliable performance in real-world settings\, from household assistance to warehouse logistics\, requires the agent to tightly integrate perception\, semantic reasoning\, and long-horizon planning under cluttered layouts\, ambiguous appearances\, and robot-specific constraints. Vision-language models (VLMs) offer rich semantic priors for this task\, but directly inserting them into the navigation loop often leads to inefficient exploration\, unstable behavior\, and limited transfer across platforms. This thesis argues that these failures stem from a multi-level misalignment between how VLMs reason and what navigation demands\, and presents four complementary contributions to address it. STRIVE shows that object navigation improves substantially when the environment is summarized as a structured graph of objects\, viewpoints\, and rooms\, letting the VLM reason at a semantic level while classical algorithms handle local exploration. SysNav extends this into a deployable system by decoupling semantic reasoning\, room-level planning\, and embodiment-specific control for robust cross-platform deployment. IntentNav shifts from prompting to learning\, showing that navigation decisions become more stable when trained with intent-aligned supervision from human demonstrations. Recognizing that object-goal search underutilizes VLM reasoning\, Goal2Pixel moves to instruction-guided navigation where longer\, compositional instructions demand richer language grounding\, and reformulates the task as pixel grounding so the model directly connects instruction understanding to executable motion. Together\, these works trace a progression from structured representation through system integration and learned decision making to instruction-guided navigation\, arguing that effective embodied navigation with VLMs requires aligning reasoning with the right representations\, architectures\, learning objectives\, and task formulations.\n \nCommittee:\nProf. Ji Zhang (Chair)\nProf. Wenshan Wang\nZhixuan Liu
URL:https://www.ri.cmu.edu/event/towards-generalizable-embodied-navigation-with-vision-language-models/
LOCATION:Newell-Simon Hall 3305
CATEGORIES:MSR Thesis Presentation,Student Talks
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20260624T140000
DTEND;TZID=America/New_York:20260624T150000
DTSTAMP:20260711T161020
CREATED:20260622T141205Z
LAST-MODIFIED:20260622T141205Z
UID:151631-1782309600-1782313200@www.ri.cmu.edu
SUMMARY:Towards Socially Intelligent Multi-Agent Systems: Zero-Shot MARL Coordination and Theory-of-Mind Benchmarking of LLM Agents for Strategic Deception
DESCRIPTION:Abstract:\nAn agent that performs well on its own may still struggle when working with others. In multi-agent environments\, success depends not only on understanding the world but also on understanding what other agents know\, intend\, and conceal. Cooperative partners follow hidden conventions\, while adversarial opponents deceive. This work argues that robust multi-agent behavior requires explicit reasoning about these hidden mental states\, and that we must measure this reasoning directly rather than simply looking at task outcomes. \nThese concepts are developed through two complementary projects. The first\, BEACON\, addresses the zero-shot coordination problem: how can an agent coordinate effectively with unfamiliar partners it has never trained with? When agents learn from offline data\, they often lock into dataset-specific conventions that work well with familiar partners but fail with new ones. BEACON is an offline-to-online learning framework that clusters offline trajectories into different conventions\, trains diverse specialists for each convention\, and uses belief-conditioned counterfactual rollouts to adapt online. On 2- and 3-player Hanabi\, BEACON achieves state-of-the-art zero-shot coordination performance while using up to five times fewer training frames than strong online baselines. It also coordinates with human partners comparably to a leading online method. The second project\, AmongUs-X\, asks whether large language model agents genuinely deceive or merely win through other means. Built on the social-deduction game Among Us and spanning 21 model families across more than 8\,700 games\, the benchmark elicits agents’ beliefs at fixed points during meetings. This yields eight Theory-of-Mind metrics measuring detection\, deception\, influence\, and grounding. Win-rate-derived ratings track crewmate detection but miss impostor deception entirely. However\, the elicited beliefs remain well-calibrated\, enabling direct mechanism-level evaluation. \nBoth projects arrive at the same conclusion: high self-play scores can hide poor coordination\, and high win rates can hide absent deception. Modeling other agents’ hidden information and measuring that modeling explicitly is essential for building socially intelligent multi-agent systems and evaluating them reliably. \nCommittee:\nDr. Katia Sycara (chair)\nDr. Jiaoyang Li\nRenos Zabounidis
URL:https://www.ri.cmu.edu/event/towards-socially-intelligent-multi-agent-systems-zero-shot-marl-coordination-and-theory-of-mind-benchmarking-of-llm-agents-for-strategic-deception/
LOCATION:Newell-Simon Hall 4305
CATEGORIES:MSR Thesis Presentation,Student Talks
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20260625T130000
DTEND;TZID=America/New_York:20260625T140000
DTSTAMP:20260711T161020
CREATED:20260622T141019Z
LAST-MODIFIED:20260622T141019Z
UID:151629-1782392400-1782396000@www.ri.cmu.edu
SUMMARY:Data-Driven Representation and Reasoning for Aviation Safety
DESCRIPTION:Abstract:\nAviation safety analysis increasingly benefits from large-scale operational trajectory data\, yet raw motion traces alone are insufficient for understanding safety-critical events on the airport surface. The significance of an aircraft’s motion depends on the structured operational environment in which it occurs\, including runways\, taxiways\, hold-short lines\, and interactions among multiple agents over time.\n\nThis thesis presents a broader framework for AI-enabled aviation safety and develops several components to support it. Amelia-42 dataset provides a large-scale operational data foundation for studying airport surface movement. Trajectory alignment with airport surface graphs recovers snapped routes\, route-transition predictions\, candidate conflict points\, and time-to-node estimates. To support interpretable safety reasoning\, World2Rules introduces a neural-symbolic pipeline that learns human-interpretable runway-incursion rules from incident reports and nominal operational observations. Critical Scenario Identification mines real runway-incursion reports and evaluates whether models can identify the critical agents and timestamps in safety-relevant interactions. Together\, these components demonstrate how trajectory data can be transformed into structured representations\, interpretable safety rules\, and evaluation methodologies that support the identification\, explanation\, and analysis of safety-critical aviation scenarios.\n\nCommittee:\nDr. Sebastian Scherer (Chair)\nDr. Jean Oh\nJunwon Seo
URL:https://www.ri.cmu.edu/event/data-driven-representation-and-reasoning-for-aviation-safety/
LOCATION:Newell-Simon Hall 4305
CATEGORIES:MSR Thesis Presentation,Student Talks
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20260625T133000
DTEND;TZID=America/New_York:20260625T143000
DTSTAMP:20260711T161020
CREATED:20260623T134347Z
LAST-MODIFIED:20260623T134347Z
UID:151646-1782394200-1782397800@www.ri.cmu.edu
SUMMARY:Brain-Aligned Tactile Representations for Dexterous Robot Learning
DESCRIPTION:Abstract: Touch is the essential sensory modality through which animals and robots physically negotiate the world. While much of robotic touch focuses on the capabilities of currently available tactile hardware\, this thesis asks a more general question: what forms of tactile processing and representation could allow robots to approach the dexterity of animals? This thesis argues that simulated force and torque provide a privileged representation of touch: one that is biologically grounded\, computationally scalable\, and effective for dexterous behavior. \n\n\n\n\nFor biological touch\, the rodent vibrissal pathway provides a tractable model system for studying tactile processing.  We train task-optimized temporal networks on realistic whisker force/torque sequences to identify shapes and compare their internal representations against recordings from rats’ barrel cortex. We find that convolutional recurrent models (ConvRNNs) align very closely with neural data\, and contrastive self-supervision with tactile-specific augmentations matches supervised alignment as a label-free proxy.\nIn robotic touch\, we test whether the same class of representation is useful for dexterous behavior. We develop a GPU-parallel tactile sensor simulator that exposes a family of tactile abstractions under one interface\, from binary contact and per-taxel force/torque to marker displacement and temperature\, fast enough to serve as the front-end for dexterous reinforcement learning. For in-hand manipulation tasks\, we find that sensor placement matters more than sensor type or resolution\, that whole-hand coverage closes most of the gap to a privileged teacher. Across 3 dexterous tasks\, per-taxel force/torque emerges as the most useful and robust observation.\nTogether\, these results argue that force and torque are not merely convenient signals to simulate\, but a biologically grounded and behaviorally effective representation of touch. With recurrence and broad spatial coverage\, simulated force and torque provide a common representational substrate for understanding tactile computation in brains and building tactile intelligence in robots.\n\nCommittee:\nProf. Aran Nayebi (Chair)\nProf. Katerina Fragkiadaki\nAkash Sharma
URL:https://www.ri.cmu.edu/event/brain-aligned-tactile-representations-for-dexterous-robot-learning/
LOCATION:Gates Hillman Center 4405
CATEGORIES:MSR Thesis Presentation,Student Talks
END:VEVENT
END:VCALENDAR