Debating the Feasibility of Scaling Large Neural Networks for Robotics: Insights from CoRL 2022
The Conference on Robot Learning (CoRL) is an annual event that brings together researchers, practitioners, and enthusiasts in the field of robotics to discuss the latest developments and trends. Last year’s CoRL was particularly notable for its focus on the debate surrounding the feasibility of training large neural networks on very large datasets to solve robotics problems. With over 900 attendees, 11 workshops, and almost 200 accepted papers, it was the biggest CoRL yet.
The debate on scaling as a solution to robotics was at the forefront of many discussions at the conference. The idea that training a large model on a massive dataset could lead to significant advancements in robotics has gained traction in recent years, especially in light of the success of large-scale models in other domains such as Computer Vision and Natural Language Processing.
Proponents of scaling in robotics argue that the success of large models in vision and language tasks suggests that a similar approach could work for robotics. They point to recent papers and projects that show promising results when training models on large robotics datasets. They believe that scaling up models could lead to significant breakthroughs in solving general robotics tasks.
However, there are also skeptics who question the practicality and effectiveness of scaling as a solution to robotics. They raise concerns about the lack of real-world data, the variety of robot embodiments, the complexity of robotics tasks, and the high cost of training large models. They argue that even if scaling works to some extent, it may not fully solve the challenges in robotics, especially when it comes to achieving high levels of accuracy and reliability.
The debate at CoRL generated a range of perspectives and insights on the topic. Some researchers suggested exploring new directions, such as leveraging simulation, combining classical and learning-based approaches, and focusing on real-world mobile manipulation tasks. Others emphasized the importance of reporting negative results and promoting ease of use in robot learning systems.
In conclusion, while the debate on scaling as a solution to robotics continues, there is consensus within the community that exploring new approaches and addressing existing challenges are vital for advancing the field. By sharing insights, collaborating on research, and being open to new ideas, researchers can work towards overcoming the limitations and maximizing the potential of large-scale models in robotics.
This post originally appeared on the author’s personal blog and reflects their experience and observations at the Conference on Robot Learning.