Can Large Neural Networks Solve Robotics? Insights from CoRL 2023
At CoRL 2023, a central debate emerged: can training large neural networks on massive datasets solve robotics? Proponents argued that the success of large models in computer vision and NLP suggests this approach is promising, citing initial results from Google DeepMind's RT-X and RT-2 as examples. They believe the ongoing advancements in data and compute power fuel this direction. However, critics pointed out the current scarcity of robotics data, the immense variability across robot embodiments and environments, and the prohibitive cost of collecting large-scale datasets. Furthermore, even achieving high accuracy might not translate to the 99.X% reliability needed for practical deployment. Some suggested combining classical control methods with learning, while others called for entirely new approaches. Ultimately, CoRL 2023 highlighted the opportunities and challenges in robotics, offering valuable insights for future research.
Read more