Learned or engineered? The split over how robots should get smart
Russ Tedrake argues for rigorous thinking in the age of deep learning, Michael Littman discusses the bitter lesson, Andrej Karpathy champions Software 2.0, and Marc Raibert wants to join athletic intelligence with cognitive intelligence. Comparing them clarifies what data-driven learning has and has not yet delivered for machines that must touch the real world.
What is the real disagreement over how robots should get smart? Show answer
Everyone uses learning. They differ on how much a person must still understand and engineer, and on how far driving and game results transfer to touching objects.
What is Andrej Karpathy's Software 2.0 idea? Show answer
More and more code is written in the weights of neural networks. Programming becomes curating datasets and objectives.
What history does Karpathy cite for learning over hand-written code? Show answer
Computer vision moved from hand-designed algorithms to networks that learn everything, and each removal of human design improved results.
What is the data engine? Show answer
A loop: deploy the model, find rare cases where it struggles, reconstruct what happened, add them to the training set and retrain.
What did the early World of Bits project teach Karpathy? Show answer
Agents trained from scratch never stumbled on success. Learning worked once it started from a pretrained model with prior knowledge.
What is Russ Tedrake's worry about deep learning? Show answer
It makes it easy to get things working without thinking critically. He values clear thinking and knowing each tool's limits.
What is Tedrake's matrix example? Show answer
We already know how to invert a matrix well. You could train a network to do it, but you probably should not.
Why is contact so hard for robots? Show answer
Making or breaking contact changes the system abruptly, and the number of possible contact configurations grows explosively.
What does Marc Raibert report about today's best robot performances? Show answer
Most, including Atlas's, still come from traditional control such as model predictive control. He expects combining it with learning to do best.
What reservation does Michael Littman have about scale winning? Show answer
Growth in computing may be the front half of an S-curve, and learned successes depend on rare human skill.
How do Karpathy and Tedrake differ on hardware? Show answer
Karpathy sees each extra sensor as a liability and thinks vision suffices. Tedrake wants softer robots with tactile sensing for contact.
Why should one success not be extrapolated? Show answer
A backgammon program's success did not generalize for decades, and a system improving at Go is only improving at Go.