Great piece, Chris. The Cambrian Explosion framing is compelling, maybe even more than you intended. Because the other half of the Cambrian story is massive extinction. Most of those phyla didn't make it. The explosion produced the diversity, but what survived was determined by something else entirely.
I believe simulation will keep getting better, and it's actually one of the fields I'm working in right now, so I'd bet the explosion happens. I'm also really curious what determines which form factors survive it.
I wrote about wizard of oz in robotics regarding the Chinese viral video. It is true we came close yet so far to control these robots.
I want to be better to write this kind of topic just like you. I really adoring your writing Chis. The topic that hard to be found and yet always fascinated me.
Not really! I think operationally indistinguishable might be possible though, i.e. some amount of targeted data collection gives you a simulation that's good enough for your problem.
This is where cross embodiment comes in, in the original essay. Otherwise it seems overly laborious to collect data for all these crazy form factors we are imagining.
Agreed that we can only start guessing at the right questions to ask, in a way. We are approaching it from a very “what is humanly” possible kind of perspective, or a hybrid thereof. At what point will agentic AI be unfeasible to test and double check? How do we build recursive cross checking and “back translation” into the testing process to accommodate and prove early models? How far is the line where we throw ourselves onto AI with enough layers of confidence-building protection and proofing?
I have been thinking about this for the past few weeks and have come to the conclusion that market forces almost dictate we figure out contact-rich simulation to scale to billions of robots.
- Industrial robots are useful not only because they replace expensive human labor, but also because they are faster, more precise and more reliable. Not quite true for humanoids (yet).
- We likely can't do online RL for tasks like manipulating delicate objects or moving heavy things around
- Current actuation technology make us choose between uncomfortable tradeoffs for humanoid performance - speed, lifting capacity, safety, weight, energy usage - that are well below the pareto frontier for humans. Motors physics is punishing in compact / small volumes compared to muscles
Great piece, Chris. The Cambrian Explosion framing is compelling, maybe even more than you intended. Because the other half of the Cambrian story is massive extinction. Most of those phyla didn't make it. The explosion produced the diversity, but what survived was determined by something else entirely.
I believe simulation will keep getting better, and it's actually one of the fields I'm working in right now, so I'd bet the explosion happens. I'm also really curious what determines which form factors survive it.
I wrote about wizard of oz in robotics regarding the Chinese viral video. It is true we came close yet so far to control these robots.
I want to be better to write this kind of topic just like you. I really adoring your writing Chis. The topic that hard to be found and yet always fascinated me.
Ya once software dev is solved the physical manifestations of it will follow and perhaps the purest expression of that is robotics in all its forms.
Simulations will continue to improve, to the point where they’re operationally indistinguishable from reality
Quite an assumption! Do you think we're on track for that?
Not really! I think operationally indistinguishable might be possible though, i.e. some amount of targeted data collection gives you a simulation that's good enough for your problem.
This is where cross embodiment comes in, in the original essay. Otherwise it seems overly laborious to collect data for all these crazy form factors we are imagining.
Agreed that we can only start guessing at the right questions to ask, in a way. We are approaching it from a very “what is humanly” possible kind of perspective, or a hybrid thereof. At what point will agentic AI be unfeasible to test and double check? How do we build recursive cross checking and “back translation” into the testing process to accommodate and prove early models? How far is the line where we throw ourselves onto AI with enough layers of confidence-building protection and proofing?
I have been thinking about this for the past few weeks and have come to the conclusion that market forces almost dictate we figure out contact-rich simulation to scale to billions of robots.
- Industrial robots are useful not only because they replace expensive human labor, but also because they are faster, more precise and more reliable. Not quite true for humanoids (yet).
- We likely can't do online RL for tasks like manipulating delicate objects or moving heavy things around
- Current actuation technology make us choose between uncomfortable tradeoffs for humanoid performance - speed, lifting capacity, safety, weight, energy usage - that are well below the pareto frontier for humans. Motors physics is punishing in compact / small volumes compared to muscles