Toward Reliable and Scalable Autonomy in Human Environments
Robot learning has made rapid progress on isolated capabilities such as dexterous manipulation, locomotion, navigation, physical reasoning, and increasingly general vision-language-action models. Yet a central premise of real-world deployment is that robots must operate in human environments and around humans: homes, hospitals, offices, warehouses, and public spaces in which people are not only end-users, but also collaborators, supervisors, demonstrators, and bystanders. These humans may introduce ambiguous instructions, changing preferences, social expectations, physical constraints, and safety-critical interruptions that current robot learning systems rarely handle in a reliable and scalable way. Translating today's lab-scale demonstrations into real-world systems that are reliable enough to be trusted, and scalable enough to be deployed across diverse users, tasks, and settings, remains an open challenge that cuts across policy learning, perception, interaction, adaptation, and safety.
This workshop convenes researchers across robot learning, human-robot interaction, foundation models, and safe autonomy to focus on the human-centered axis of robot development and deployment. We are interested in innovative approaches that treat humans as a first-class component of the learning, adaptation, evaluation, and deployment loop: how robots learn from people at scale, interact with them safely and intuitively, adapt to individual preferences and physical contexts, and recover gracefully when assumptions about the world or the human break.
The workshop centers on six open problems. We welcome submissions addressing any of these problems, as well as other closely related research questions.
How do we collect interaction data — teleoperation, egocentric video, language feedback, in-the-wild traces — at the scale and quality required for generalist policies in human environments, while respecting privacy, consent, and human-effort budgets?
What benchmarks and protocols capture safety margins, social appropriateness, recovery from failure, and long-horizon robustness when a human is part of the closed loop, rather than only static task-completion rates on fixed scenes?
How can vision-language-action models be grounded in physical contact, social norms, and human attention so that their outputs respect what people intend and expect, instead of optimizing for object-level affordances in isolation?
What does it take to keep policies stable in the face of clutter, partial or noisy demonstrations, interruptions, ambiguous instructions, and adversarial or untrained users, particularly when the same robot is deployed across many homes or facilities?
How should robots reason about turn-taking, intent inference, and repair dialogues over minutes-to-hours of shared activity with humans, rather than single-shot commands followed by open-loop execution?
How can deployed robots learn online from preferences, corrections, and language feedback without catastrophic forgetting, unsafe exploration, or unbounded annotation cost — and how do we audit what they have learned from each human user?
National University of Singapore
Associate Professor at the National University of Singapore, where he leads the Collaborative, Learning, and Adaptive Robots (CLeAR) group. His work builds trustworthy interactive robots, spanning trust-aware human-robot interaction, tactile perception, and learning robot policies from human supervision and feedback.
Massachusetts Institute of Technology
Assistant Professor at MIT (AeroAstro and CSAIL). Her research studies how robots can learn the right representations of human intent, combining robot learning, algorithmic human-robot interaction, and learning from human feedback to keep autonomous systems aligned with people's preferences.
Stanford University
Postdoctoral scholar in the Autonomous Systems Lab at Stanford University. He works on safe human-robot collaboration, combining formal methods such as set-based reachability analysis with reinforcement learning to build provably safe control shields, and on rapidly personalizing robot behavior to individual users.
Toyota Research Institute
Research Lead and Manager at the Toyota Research Institute, where she co-leads the Trustworthy Learning under Uncertainty direction. Her work spans statistical robot evaluation, failure detection and recovery, and interactive learning for large behavior models, after earlier work on uncertainty-aware perception for autonomous driving.
Opening & live challenge poll
Invited talks, clustered
Four 20-min talks + Q&A in two themed clusters with cross-talk.
Poster & live-demo session
Fishbowl discussion
Contributed spotlight talks
Closing synthesis & roadmap
We welcome in-progress work, negative results, system and demo papers, and cross-disciplinary contributions across any of the six challenges above.
This is a non-archival workshop: accepted papers will be posted on this website and will not appear in any proceedings, leaving you free to submit the work elsewhere. We plan to accept both published and unpublished work, including papers currently under review at other venues.
Submissions must use the official CoRL template and are limited to 8 pages of main text, excluding references and appendices. Papers will undergo double-blind review through OpenReview. All accepted papers will be presented as posters, with 3–4 papers selected for spotlight talks. We will also present a Best Paper Award.
Georgia Institute of Technology
Stanford University
StarWith
Georgia Institute of Technology
Tsinghua University
Texas A&M University
Carnegie Mellon University
Toyota InfoTech Labs
Toyota InfoTech Labs
Toyota InfoTech Labs
Flexion Robotics
Flexion Robotics
