A basic problem with getting to deep learning is that in control theoretic terms, deep learning is feedforward control, not feedback. The heavy lift is offline training, and the online inference step is only “feedback” in a very limited way, barely more than a lookup table.
Conversation
Here, even the simplest maneuvers would have 1-2 parameters that would need to be filled in with reference to live video. Eg, a “jump” modeled as a parabolic arc is 3 parameters to lock down using a stereo inage. Not like picking the next word in a language model.
1
10
So, heavier, more expressive inference steps in online feedback part of solution means more compute. Which will be available in the next few years. Theoretical advances also needed, for more expressive quantified “languages” with stricter “grammars”
1
6
Btw when I say non-defensible, I mean against a reasonably funded competitor. Like an American startup with ~30m in funding or a Chinese state lab, and 2 years time. 90% of that would be software and production scaling costs. In 10y hobbyists will be beating these.
2
18
Never lose sight of the fact that imitation costs ~70% what innovation dies, with much lower risk if failure, and that sector pioneers usually end up with ~7% of a market once it’s mature. See Oded Shenker book, Copycats.
1
2
39
For a pioneer, maintaining a 1-2 year lead and a corresponding margin premium is the only way to stay in the game, since they usually won’t win the volumes. The factors that make them good at pioneering will make them bad at volumes. Hence the moat importance. I hope BD find one.
2
2
22
One possible moat is in social networking the robots into intelligent multiagent swarms. Not marching in parades or flash mobs. Remember beauty = symmetry = low info = no moat). More like a team of say 50 robots peer-learning a work environment and forming a work-mesh in it.
1
18
Like say a robotic wildfire fighting fire department. A handful of experienced humans leading a hundred robots. Stop using poorly paid prison inmates without better options. Owning the swarm-level software orchestration capability = you’re the next Google.
4
1
19
Speaking of which, at last night’s rover meeting, we had a talk by Victor Hill on using Docker and Kubernetes to orchestrate the rovers we’re building. He’s already got a small RPi based wheeled rover hooked up that way to a cloud server via VPN. Video soon 😎
1
10
Exciting times for robotics entrepreneurship. Somewhere on the Zoom side-channel for a boring online lecture, the Jobs/Wozniak or Page/Brin of robotics are chatting about dropping out of college and building the Apple or Google of robotics.
1
15
BD is more like SRI, doing the Doug Engelbart mother-of-all-demos equivalent. Or like PARC. There are sadly more ways for them to fumble this future than win it, but the good news is *somebody* or many somebodies, will win.
Replying to
The metaverse will run on robotic hardware at the network edge, and it will be a fat edge. The Internet of Things will become the Internet of Robots soon.
4
1
8
I doubt it. There’s probably far less useful patent coverage than we think. Math can’t be patented. Patent based IP strategy works much better in fields like chemicals, materials, or novel uses of physics. And there will be plenty of competing IP to fuel a cross-licensing swamp.
Quote Tweet
Replying to @vgr
Why is the moat not IP lawsuits? I'm not suggesting that as a normative prerogative, but rather a de facto component of the space. Is that not enough?
2
5
For calibration, when I was at Xerox, the record-setting patent winners were mostly from the toner chemistry and xerography physics side of the labs, even though hardware and software parts were nearly as complex. With robotics, I’d say 90% of important knowledge is public domain
3
14
Doh. Should have looked at specs page. They basically confirm my hypothesis in outline form.
1
5
I really should Google and search more before trying to guess things. Showing both my age and guess culture > ask culture bias here. When I see something I guess at how it might work rather than ask people or Google.
2
16

