Learned or engineered? The split over how robots should get smart
October 6, 2026 · 19 min read
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.
The rules, numbers and steps worth keeping in front of you.
- All four use learning; the dispute is how much humans must still understand and engineer.
- Karpathy: more of the software should be learned; curate data and objectives instead of writing code.
- Karpathy's data engine: deploy, find rare failures, reconstruct the truth, add to training, retrain.
- Karpathy: learning from scratch failed; starting from a pretrained model worked.
- Karpathy: every extra sensor is a liability; best part is no part.
- Tedrake: deep learning makes it easy to get things working without thinking critically.
- Tedrake: do not train a network for what mature methods already do well.
- Tedrake: contact is discontinuous and combinatorial, the hard problem for robots.
- Tedrake: hunt for rare failures and learn across fleets, since one robot cannot see enough.
- Raibert: the best physical performances so far come from traditional predictive control.
- Raibert: expect the best results from combining engineered control with learning.
- Littman: scale beat cleverness, but the growth may be an S-curve.
- Littman: learned systems succeed in the hands of skilled people; one success may not generalize.
- Driving is not manipulation; success in one does not settle the other.
This is the short form. The full insight has every expert's reasoning, their own words and the sources.