CoRL 2024 Day 3 17:30-19:00
Conference of Robot Learning keynote speakers and audience members discuss scaling, priors, robot hardware, evaluation, and what it would mean to solve robotics at CoRL 2024.
In this talk
- 10:02
Scaling is only part of the work
Panelists disagree on whether robotics can advance by scaling data and models alone. They argue that researchers need to specify what they are scaling, improve data quality, and keep multiple research approaches active. One speaker points to infrastructure and social acceptance as conditions for robots to thrive.
- 16:47
Priors, rewards, and compositionality
The discussion turns to what useful priors look like and whether behavior can emerge from data. Panelists highlight compositionality, curricula, reward design, and the structure of bodies and learning systems as open questions. They also debate whether hierarchy and composition need to be explicit or can emerge inside large models.
- 32:44
Hardware, long-tail cases, and benchmarks
One panelist says current hardware is capable enough to expose shortcomings in methods, while robust tactile sensing remains a concern. On long-tail failures, panelists say robotics still needs to solve common tasks such as moving objects to reach a cup. They favor honest evaluation, clear metrics, baselines, and reproducible reporting, while expressing doubts about a single global benchmark.
- 53:04
Defining progress and robotics solved
Panelists offer different milestones for robotics, including child-level dexterity, a useful home robot, and systems that can be built reliably outside a specialist lab. They describe practical goals such as cleaning, cooking, and support in elder care, alongside questions of safety and social acceptance. They compare the field's evaluation maturity with stages in computer vision, while noting that robotics lacks comparable passive datasets.
- 1:03:54
AI, research communities, and conference close
The panel discusses robot consciousness, AGI claims, human-designed priors, and overlap between robotics, machine learning, and computer vision communities. After the discussion, the conference announces its paper awards and that CoRL 2025 will take place in Seoul from September 27 to 30, co-located with Humanoids 2025.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.