HackerA driver took over. What should a driving system learn?
One intersection, a takeover log and a software update. Accounts from openpilot, Tesla and Waymo help separate clues to a failure, examples worth training on and old scenarios that can test a fix.
Questions in this reading · 4
What is the stationary return at the light?
Approaching a red light, the radar reports a stationary object. Braking for a sign might be an error. Continuing toward the back of a queue might also be an error. Hotz uses those opposite consequences to explain a problem openpilot faced at the time. Write down what both explanations predict, then ask which observation separates them. He points to vision. A log entry saying only that the driver braked loses precisely the information needed to diagnose this case.
Passages and timestamps · 2
it's hard to differentiate a stopped car from a signpostYou have to do this visuallyWhat does a single takeover label?
A driver nudges the wheel at an intersection. Musk offers two explanations for the same action: convenience, or a mistake by the system. In a different interview Hotz sketches a loop that feeds takeover moments back into training. Together they specify a labeling task. Recover the intended path, mark what the driver changed and find the detail in the scene that justified the change. A takeover count alone cannot separate the two motives.
Passages and timestamps · 4
did they take over for conveniencewhat is the optimal spline for traversing an intersectionuser disengages, we train on thatwith some labeling could indicate the failure of the systemFind more cases, then put the revision back at the intersection
Once a stopped car is identified as the failure mode, keep two lists. Karpathy wants more examples in a poorly covered part of the training set. Arnoud describes putting revised software back into recorded scenarios alongside road driving and deliberately constructed tests. For the same intersection, write down which clips will teach the model and which conditions will test the change. Improvement on an intersection used in training leaves the question open at another one.
Passages and timestamps · 4
they weren't told what to do in those rare casesmine the parts that are not yet represented well in the data setanother one is around simulationreplay all those miles that you've drivenHow far can a staged obstacle test a driving system?
The presenters use dummies and disguised scenery to observe whether vehicles stop. Log the active system, target position and response rather than reducing the run to a sensor winner. To turn it into a regression test, fix the target, speed, path and pass condition. A constructed encounter can supply a reproducible failure lead; a ranking of sensing strategies requires broader coverage.
Passages and timestamps · 2
only relies on simple cameras and its image processingdetect the kidTry it on your own example
Take one human intervention. Write down what the system seemed about to do, what the human changed and the environmental cue that justified it. Reserve a different intersection. What should the revision do there before you say it learned?