BAIR Brings 33 Papers to CoRL as Robotics Converges on Memory, Touch and Humanoids

Berkeley AI Research has published its schedule for CoRL 2026 in Austin. From November 9 to 12, its members will present 33 main conference papers, 4 late breaking results and 10 oral or spotlight contributions. Professor Negar Mehr will also give an Early Career Keynote on robot interaction and coordination.

Berkeley brings a broad cross section rather than one star result

The program spans memory for vision language action models, navigation, behavior cloning, reward learning, tactile manipulation, humanoid motion and multi agent systems. Many projects address operation beyond a short laboratory demo, including asynchronous VLA navigation on edge hardware, safety filters and whole body humanoid control.

The count covers papers with at least one UC Berkeley author, including joint affiliations. The blog also lists selected workshops and program roles but excludes the main conference demo track. Its authors note that the late breaking poster list may still grow.

A common problem is transferring learned behavior into an awkward world

Across the topics, robots must preserve context beyond one action, react to physical contact and handle situations absent from training. MEM studies multi scale memory, EgoScale uses diverse egocentric human data and ResSafe filters unsafe humanoid movements.

For deployment teams, that is more useful than another isolated benchmark. Robotics is becoming a stack of perception, planning, control, safety and adaptation that must work together. Berkeley's broad program suggests that no single model closes that stack yet.

Conference selection does not show what survives outside the lab

CoRL 2026 lists 687 accepted papers overall, including 21 oral and 72 spotlight papers. One institution's strong presence shows research breadth, not automatic superiority of any method. The schedule also offers no common comparison of success rates, training data or hardware cost.

Experimental detail matters especially in robotics. A video of one successful run does not reveal failures, operator interventions or sensitivity to a new environment. Honest assessment requires the individual papers, evals and released code, not program titles alone.

Replication and long runs will separate methods from conference demos

After the conference, watch which teams release data, code and exact protocols. Independent replications on a different robot, in another room and across repeated runs will provide stronger evidence.

Integration matters just as much. Memory without safe control and touch without reliable planning remain machine parts. CoRL will identify candidates. Sustained operation beyond a poster session will determine their value.

Lilith's verdict

A humanoid can climb a ladder once on a conference poster. Research becomes technology when it climbs down, repeats the feat 100 times and nobody has to hold a mattress underneath.