{"response":{"award":[{"abstractText":"For a robot to work safely alongside people, engineers today must manually define every dangerous situation the robot might encounter and write rules to avoid each one. This is fundamentally limited: no engineer can anticipate every hazard, the resulting rules break when\r\nconditions change, and different people have different expectations for how cautiously a robot should behave. This project starts from a different premise: the definition of safe behavior is already contained, implicitly, in the expert demonstrations that are now routinely collected for robot learning. By aggregating these demonstrations across a wide variety of tasks and operating conditions, it is possible to infer what generally safe actions look like and build a reusable safety filter — a module that composes with any robot controller to correct unsafe actions before they are executed, in settings from assembly lines to assistive devices. The filter can also be conditioned on verbal feedback from a human coworker, allowing the robot to adjust to individual preferences. The project produces open-source software and benchmark tasks for the broader research community. Educational activities include integration of project themes into graduate and undergraduate robotics courses at Cornell University and the University of Texas at Austin, and undergraduate research mentorship through summer programs. K-12 outreach activities for the project include curriculum development with local school districts, and partnerships with local science centers and robotics teams in Ithaca and Austin.\r\n\r\nThis project develops a new framework for robot safety that replaces hand-crafted safety constraints with learned, implicit representations of safe behavior. Existing safety methods in robotics require explicit mathematical descriptions of which states are safe, but these descriptions are difficult to specify, computationally expensive to enforce, and brittle in novel situations. Even recent learned approaches that relax the need for hand-designed constraints still require training multiple tightly coupled models and collecting demonstrations of unsafe behavior. This project takes a more direct approach: training a denoising diffusion model on safe expert demonstrations aggregated across tasks and using it as a safety filter that projects arbitrary robot actions back toward the demonstrated distribution. First, the research team formalizes the connection between diffusion-based denoising and safety filtering, establishing probabilistic safety guarantees grounded in stochastic process theory and learned safety critics; in tabletop manipulation tasks such as pick-and-place near fragile objects, empirical failure rates are compared against these bounds using only camera observations, without access to ground-truth state. Second, the project develops methods for incorporating human feedback into the safety filter, including natural language commands such as \"slow down near the glass.” Third, the project introduces techniques for detecting situations outside the training distribution and triggering a recovery policy before failure occurs, such as when a camera degrades or a motor joint locks. The research team evaluates all three components through physical robot experiments and user studies spanning manipulation and navigation tasks, with quantitative comparisons to classical and learned safety filters.\r\n\r\nThis award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.","activeAwd":"true","agency":"NSF","awardAgencyCode":"4900","awardee":"UNIVERSITY OF TEXAS AT AUSTIN","awardeeAddress":"110 INNER CAMPUS DR","awardeeCity":"AUSTIN","awardeeCountryCode":"US","awardeeDistrict":"25","awardeeDistrictCode":"TX25","awardeeName":"University of Texas at Austin","awardeePhone":"5124716424","awardeeStateCode":"TX","awardeeZipCode":"787121139","cfdaNumber":"47.041","date":"08/13/2026","dirAbbr":"ENG","divAbbr":"ECCS","estimatedTotalAmt":"360000","expDate":"09/30/2029","fundAgencyCode":"4900","fundProgramName":"FRR-Foundationl Rsrch Robotics","fundsObligated":["FY 2026 = $360,000.00"],"fundsObligatedAmt":"360000","histAwd":"false","id":"2535276","initAmendmentDate":"08/13/2026","latestAmendmentDate":"08/13/2026","managingPec":"144Y00","orgCodeDir":"07000000","orgCodeDiv":"07010000","orgLongName":"Directorate for Engineering","orgLongName2":"Division of Electrical, Communications and Cyber Systems","orgUrl":"http://www.nsf.gov/div/index.jsp?div=eccs","parentUeiNumber":"X5NKD2NFF2V3","pdPIName":"David Fridovich-Keil","perfAddress":"110 INNER CAMPUS DR","perfCity":"AUSTIN","perfCountryCode":"US","perfDistrict":"25","perfDistrictCode":"TX25","perfLocation":"University of Texas at Austin","perfStateCode":"TX","perfZipCode":"787121139","pi":["David Fridovich-Keil dfk@utexas.edu"],"piEmail":"dfk@utexas.edu","piFirstName":"David","piId":"270061116","piLastName":"Fridovich-Keil","poEmail":"cvanwyne@nsf.gov","poName":"Cassie VanWynen","poPhone":"7032928496","primaryProgram":["01002627DB NSF RESEARCH & RELATED ACTIVIT"],"progEleCode":"144Y00","program":"Artificial Intelligence (AI), ROBOTICS","progRefCode":"075Z, 6840","publicAccessMandate":"1","startDate":"10/01/2026","title":"Collaborative Research: Robot Safety with Deep Generative Modeling","transType":"Standard Grant","ueiNumber":"V6AFQPN18437"}],"metadata":{"offset":0,"rpp":25,"totalCount":1}}}