What robotics can learn from insects

Self driving cars are developing in similar fashion like vertebrates evolved intelligence 380 Million years ago

We are in the midst of the transition from analog to digital intelligence. This is one of those big transitions, like for example when vertebrates invaded land from water. Big transitions in evolution are exciting. Even more exciting is how similar they are.

Malcom Maciver just published a paper on his thesis on how vertebrates invaded land from water and developed brain circuitry that allowed for situational planning. This is certainly a higher form of intelligence than the habit based planning of the invertebrates that inhabited land before that.

Maciver’s paper argues that invertebrates like insects lived on land before the invasion of vertebrates but they could not act on situations because they didn’t have the visual system that is required to do that. What they had was a memory of their habitat which allowed them to live according to habits. Your bee hive is there and the flowers are here and you go back and forth according to your memory. A change in the system would confuse you and make you loose your orientation. You can’t react on situational changes.

Enter vertebrates. They invaded land from water and evolved using their already developed visual system much more effectively because on land vision works much better than in water. You can see further. Thanks to farther vision vertebrates could see stuff further ahead. They developed brain circuitry that allowed for situational planning.

Maciver makes another interesting distinction. This planning is most useful in an environment with obstacles. So a desert doesn’t give you much advantage. But too many obstacles like in a jungle also don’t work. You need some intermediary environment like the Savannah in East Africa for example.

Now let’s fast forward to today. We are witnessing the transition from analog to digital intelligence. It has long been our thesis (see our manifesto) that self driving cars are the catalyst for this transition. They are the first robots that interact with humans in an uncontrolled terrestrial environment at scale.

And the transition is very similar to the vertebrate invasion 380 Million years ago. Here is a snipped from Andrej Karpathy’s presentation at CVPR’20 in June. Check the snipped about where he compares the different approaches of Waymo and Tesla. While Tesla relies solely on vision, Waymo relies on high precision HD maps. Waymo memorizes the environment and acts according to this memory. If something changes, the car cannot react appropriately.

By contrast Tesla is building a vision only AI which should be able to react on any situation. That’s the idea. Karpathy admits that this is harder. Very much like the vertebrates had to developed special brain circuitry to enable situational planing, so Tesla has to develop special AI networks to deal with situational awareness and planning.

Tesla’s problem is hard but worth pursuing.

Finally, I am fascinated by Maciver’s thesis that vertebrates needed a special kind of environment to develop situational planning. Deserts and jungles didn’t do it. The Savannah did.

Similarly, Tesla needs to find the appropriate environment to evolve. Highways are too easy, inner city streets too tough. They need a middle way, like the Savannah. What could that be? For example you daily commute. Or the trip to your weekend house. Something where complexity is high but not chaotic.