Nitin Sitaraman

Architecting nervous systems for humanoids at-scale | Physical AI, Robotics and Sensors🤖

[Nitin Sitaraman @nitinsitaraman]

@elonmusk The end of tokenmaxxing and return of common sense. How to use ai tools effectively to launch individuals and firms into the stratosphere is going to be key for the winners of the next decade

[Nitin Sitaraman @nitinsitaraman]

This is a collar, a well known option hedging strategy. If the stock popped as he claimed then he actually lost money as he wouldn’t make any money as proved moved above his call strike price. Anyone can do this trade in a brokerage account. Nice to know Mark doesn’t know what he’s talking about.

[Nitin Sitaraman @nitinsitaraman]

@norse_creative @ICannot_Enough @TheHumanoidHub First time I’ve seen such a nuanced opinion on product development. I agree with the general sentiment

[Nitin Sitaraman @nitinsitaraman]

@CyberRobooo NEO has benefits due to it's tendon system but what about the disadvantages? By going fully tendon driven what tradeoffs have been made on NEO due to this architecture?

[Nitin Sitaraman @nitinsitaraman]

The #1 silent killer destroying humanoid perception in the real world — and almost nobody talks about it.
High-power actuator and distribution noise wrecks sensor signal integrity before your AI models even have a chance.
I break down the full end-to-end solution in this deep dive.
If you’re building real humanoids, this is required reading.
https://t.co/qhO8pJkdLt
Comment or DM me on the toughest humanoid robotics hardware design or integration challenge you’ve faced.

[Nitin Sitaraman @nitinsitaraman]

https://t.co/SxGP39D9rM

[Nitin Sitaraman @nitinsitaraman]

https://t.co/SxGP39D9rM

[Nitin Sitaraman @nitinsitaraman]

Given we are still waiting after decades of video game development for a high fidelity high throughput physics simulator I expect more focus to be placed on the first approach in the next 1-3 years. This should help drive new developments in material science, reliability and engineering physics. #Robotics #PhysicalAI

[Nitin Sitaraman @nitinsitaraman]

Most companies in AI and Robotics are banking on one of two paths:

1. Build a robot and train it in the real world using real world collected data
2. Simulate the robot in a simulated environment and transfer the learnings to real, hopefully in zero-shot

#Robotics #PhysicalAI

[Nitin Sitaraman @nitinsitaraman]

The 2nd approach depends on simulating the robot and real world accurately. Physics needs to be properly captured so that both robot kinematics and dynamics are truly representative. In practice discrepancies from sim to real are common and not easy to overcome without randomization and overfitting to sparse perception datasets.

[Nitin Sitaraman @nitinsitaraman]

@elonmusk Wet physics looks off

[Nitin Sitaraman @nitinsitaraman]

@yacineMTB People get compensated highly for doing dumb work all the time. The only reason sw pays more than ee is because most people live easy lives in the digital world. Making real things operate in the real world using physics is hard. And most people don’t like doing hard things

[Nitin Sitaraman @nitinsitaraman]

@jacobrintamaki Please add me

[Nitin Sitaraman @nitinsitaraman]

While @sundayrobotics execution is commendable note that training an autonomous simple dof claw grip robot has been known since mobile aloha years ago.

Having humans do tasks intentionally with reduced dofs in a nonhuman way limits the long term capability and possibility of any humanoid bot.

This looks like an effort to preoptimize hw to sidestep the complex challenge of training a truly dextrous neural net.

I expect the wrong lessons will be learned from this.

[Nitin Sitaraman @nitinsitaraman]

@KeanuReeveals @lukas_m_ziegler Yes, it’s currently a bottleneck for humanoid hands but there’s a clearer path towards data collection and scalability

[Nitin Sitaraman @nitinsitaraman]

@rohanpaul_ai These comparisons are nonsensical given one has been architected via evolution to work on a mobile system while the other has been naively engineered to just deliver brute force compute. So a comparison is pointless

[Nitin Sitaraman @nitinsitaraman]

My point is all sensors are expensive when miniaturized into an aggressive form factor. Calling out tactile sensors specifically then results in drawing the wrong conclusions.

I do agree that usefulness can be likely even with no sensing at all in the hand. Many people make do with prosthetics with no haptic feedback