Developing general-purpose robots capable of autonomously performing diverse tasks in unstructured environments remains a central challenge in robotics. Recent demonstration-driven approaches, including learning from teleoperated robot demonstrations and human video demonstrations, have provided robots with strong behavioral priors and capable initial policies. However, imitation alone is unlikely to close the gap between broad initial competence and reliable performance across the long tail of real-world tasks, objects, and environments. Robots must improve beyond their demonstrated experience: autonomously collecting data, discovering failures, adapting to new situations, and refining their skills through interaction.
This workshop advocates for an era of experiences in robot learning, where robots continually expand their capabilities by interacting with the world. We aim to bring together researchers across robot learning, reinforcement learning, and continual learning to define the algorithmic and systems-level foundations for self-improving robots.
University of Washington
Physical Intelligence
UBC & Recursive
Dyna
We invite contributions that introduce theoretical frameworks, experimental findings, or applications that advance autonomous robot learning and self-improvement.
Robot self-improvement, continual robot learning, RL for robotics, autonomous data collection, failure discovery, safe exploration, world models, in-context robot learning, real-to-sim-to-real transfer, automatic reset systems, and reliable evaluation environments.
Submissions will be hosted on OpenReview. Each submission will receive at least two reviews and will be evaluated based on quality, relevance to the workshop, and suitability for spotlight or poster presentation.
Accepted papers will be presented through student spotlights and posters. We plan to give a best paper award and invite participants to contribute to the post-workshop artifact or white paper.
The workshop will combine invited talks, student spotlights, poster presentations, structured breakout sessions, and moderated group-wide discussions. The format is designed to actively engage participants rather than present a passive sequence of talks.
Participants will define what it means for a robot to self-improve, including assumptions about initial policy capabilities, available feedback, deployment constraints, allowed human intervention, and simulation benchmark designs.
Using a multi-purpose warehouse-robot case study, groups will design an end-to-end pipeline for autonomous long-term deployment, including data collection, failure detection, policy updates, safety, transfer, and evaluation.
Breakout groups will use shared online documents to record assumptions, proposed methods, benchmark ideas, open questions, and disagreements, then report back for moderated debate.
UT Austin & Sony AI
UT Austin
Google DeepMind
