BAIR at CoRL 2026
The Conference on Robot Learning (CoRL) runs November 9–12, 2026, in Austin, Texas, and Berkeley AI Research has published its schedule of participation. The listing counts 33 main-conference papers, four late-breaking results posters, and 10 oral and spotlight presentations split into three orals and seven spotlights, spread across four days of programming. Papers qualify when at least one author is affiliated with UC Berkeley, with joint affiliations counting. Faculty talks, session-chair duties and workshop roles are also included in the roundup. All times are Central Standard Time, UTC−6.
The main conference occupies November 9–11, with workshops on November 12. Oral and spotlight entries are listed as session windows rather than fixed presentation times, and each paper appears twice: once in its talk session and again in a poster session. Talks on the Hilton Austin stage will be streamed to viewing rooms at the JW Marriott, where the poster sessions are held. The Monday keynote block includes an Early Career Keynote II from Professor Negar Mehr, titled How Can Robots Learn to Interact and Coordinate with Others, chaired by Ken Goldberg.
Monday opens with a spotlight session on generalist policies, memory and transfer. Among the papers is MEM: Multi-Scale Embodied Memory for Vision Language Action Models, with a long author list spanning Marcel Torne, Karl Pertsch, Homer Walke, Sergey Levine, Chelsea Finn, Danny Driess and others; it reappears as Poster 42 in the evening session. Also in that spotlight is Diagnosing Compositional Generalization in Sequential Robot Tasks, with Yixiao Wang, Masayoshi Tomizuka and colleagues, posted as Poster 115. Three late-breaking agentic robotics posters are listed for Monday, including Scaling Online Exploration for Robot Agents and Flip Anything with Agentic Robotics Self-Learning from Shuangyu Xie, Kaiyuan Chen and Ken Goldberg.
Monday's evening Poster Session I adds a broad set: Large Video Planner Enables Generalizable Robot Control, How to Peel with a Knife: Aligning Fine-Grained Manipulation with Human Preference, HiRE: Hindsight Reward Editing for Policy Finetuning, Scalable Behavior Cloning with Open Data, Training, and Evaluation, AsyncVLA: An Asynchronous VLA for Fast and Robust Navigation on the Edge, GLST for online chained-goal navigation, ENPIRE on agentic robot policy self-improvement in the real world, and Energy-Based Policy Orchestration.
Tuesday's oral sessions cover reinforcement learning and motion reasoning, dexterous learning and assembly, and humanoid data, control and skills, the last chaired by Koushil Sreenath. Papers include Flow Policy Gradients for Robot Control, EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data, and SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model. A late-breaking block on Real2Sim2Real features RoboSimStudio, an interactive toolkit for rapid real-to-sim conversion of manipulation tasks, from Kaiyuan Chen, Shuangyu Xie, Ken Goldberg and collaborators including Erwin Coumans and Miles Macklin. Tuesday's poster session adds work on discrete diffusion policies as asynchronous executors, flow reversal steering for generalist policies, action chunking in behavioral cloning, tactile foundation policies (FTP-1), privileged tactile distillation for sim-to-real (PTLD), counterfactual video generation for humanoid loco-manipulation, VLK for learning from synthetic interactions in reconstructed scenes, and Social Humanoid.
Wednesday brings 12 main-conference papers, led by a spotlight session on optimization, agentic planning and evaluation chaired by Negar Mehr, featuring Playful Agentic Robot Learning and GaP, a graph-as-policy multi-agent self-learning harness for variational automation tasks. A humanoid locomotion and whole-body control spotlight follows with LadderMan on perceptive ladder climbing, ResSafe on safety filtering via residual reinforcement learning, and TANGO on navigation in cluttered environments with a whole-body vision-language-action model. The third poster session includes T-Rex on tactile-reactive dexterous manipulation, a very large collaboration across Berkeley and industry researchers, plus tactile curiosity work, robust perceptive locomotion (RPL), TEXEDO on test-time scaling for humanoid motion generation, CLIFT on iteratively fine-tuning Gemini Robotics on-device into humanoid specialists, and WARP-RM, a warp-augmented relative progress reward model for data curation.
Thursday's workshops include talks and organizing roles. Sergey Levine's lab speaks in the keynote block of the Roadmap to Sample-Efficient Real-World Reinforcement Learning workshop. Jitendra Malik gives an invited talk at Bimanual Dexterous Tool Use, where T-Rex and Do as I Do appear as spotlights and in the poster and demo session. Ken Goldberg is an invited speaker at Agentic Robotics, which also hosts live demos of ASPIRE, Agentic World Action Models and Reconstruct, Practice, Go Real. Claire Tomlin speaks in a shared invited block with Michael Lutter of Boston Dynamics at Sim-to-Real-to-Field. Berkeley people also organize workshops: Lasse Peters and Negar Mehr's lab on Do Robots Need World Models?, S. Shankar Sastry and colleagues on Sim-to-Real-to-Field, Sergey Levine as faculty advisor plus Paul Zhou and Renhao Wang on a workshop titled Oops, I Erred, and Jitendra Malik on Dexterous Manipulation: Full-Stack Robotics on Learning and Design, scheduled 2:00–6:00 PM.
A few caveats come with the listing: workshop schedules may shift, poster numbers follow an unofficial paper index and should be checked against the official program on the day, and the main-conference demo track is not covered. Affiliations were reviewed November 6, 2026, with late-breaking results updated November 8.
Reading the slate, several clusters stand out for developers. One is memory and long-horizon competence in vision-language-action models, where MEM's multi-scale embodied memory and the action-chunking and flow-steering papers all probe why generalist policies succeed or fail over extended tasks. Another is data: egocentric human video, synthetic interactions, counterfactual generation and open behavior-cloning datasets all point at reducing reliance on costly robot teleoperation. A third is humanoids, where locomotion, safety filtering and whole-body manipulation appear repeatedly. Tactile sensing rounds out the picture, with several papers treating contact as a first-class signal rather than an afterthought. The emphasis on self-improvement and agentic scaffolding, from ENPIRE to ASPIRE, suggests the field is moving toward systems that generate their own training signal, though the workshop-format results are early and should be read accordingly.
Why it matters: Teams building robot learning stacks get a concentrated view of where Berkeley-affiliated research is heading, particularly around embodied memory, humanoid control, tactile manipulation and data scaling. Several papers, such as RoboSimStudio and open behavior-cloning work, appear aimed at reusable tooling rather than one-off demos, which could shorten the path from simulation to deployment. Because many results are released alongside the conference and the schedule includes live demos, practitioners can expect fresh code and project pages in the coming weeks; that expectation is inference on my part, not something the schedule states.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
The Conference on Robot Learning (CoRL) takes place November 9–12, 2026, in Austin, Texas. BAIR faculty, postdocs and students will be presenting research at various sessions, including an Early Career Keynote from Professor Negar Mehr. See below for the full schedule of sessions with BAIR members, including papers co-authored with collaborators at other universities and companies. November 9–12, 2026 Austin, Texas All times CST · UTC−6 33 Main-conference papers 4 Late-breaking results (LBR) posters 10 Orals & spot