The Autonomous Future: Humanoids vs. Self-Driving Cars - Part 2: Human(oid)s have it easier...

This is part 2 of a 3 part series of post on the topic. You can find part 1 and part 3 here.

So, why would power, compute, and cooling-constrained humanoids beat robo-taxis to market? Let's examine three key advantages compared to autonomous vehicles:

1)  Interacting at Human Speed: Most human interactions, insofar as they don't involve driving a car or flying a fighter jet, but concern a bounded set of tasks (that these humanoid machines could initially be useful for) such as domestic chores, factory work, or running errands, happen at a relatively docile pace. You do not need your humanoid to sprint through the supermarket aisles, weaving through human shoppers at 30mph and grabbing produce with sub-millisecond precision from shelves and crates to be useful. If the bot can do it at a similar pace to a human, then it is already very useful. Thus, the decision window becomes larger compared to a 2.5-tonne object hurtling down a freeway at 85 miles per hour. If our shopper bot were to have some moments of lag, it might seem to be standing puzzled in front of a shelf of canned beans; if our autonomous vehicle experiences lag, it might rear-end an emergency vehicle parked on the hard shoulder. The stakes are higher and the inference window is narrower in an automotive context. Which neatly brings us to the next issue.

2)  More Benign Failure Modes: If we stick to the dominant applications envisaged for such first-generation humanoids—chores and manual labor—in many if not most circumstances, a robot encountering an intractable problem can usually safely abort and retreat from the mission in such a way as to not cause any harm to others or itself. A worker bot can step back from the machine or turn said machine off; a shopper bot can step aside and let other customers access the shelf of canned beans. Whereas a malfunctioning autonomous vehicle, in the best case, becomes an obstruction or traffic jam generator to others, and at worst, a deadly 2.5-tonne projectile obliterating whatever stands in its path. Because the speed and stakes for a vast array of useful humanoid applications are much lower and slower, humanoids actually gain more latitude with regards to the level of performance of both their compute hardware and inference algorithms.

In the next post we'll look at the third and final, and perhaps most decisive advantage: context dependency and orchestration. This is the key insight into how less powerful hardware might enable autonomous humanoids to effectively perform complex tasks, where self-driving cars often struggle.