2: The year Robotics finds the starting point

Trad (Robot) Wivery

The jaw-dropping moment of the week came courtesy of the Stanford AI Lab Professor Chelsea Finn’s grad students Tony Zhao and Zipeng Fu, with a mobile robot that could:

• cook and serve shrimp

• call and board an elevator

• wipe spilled wine

While at first causing consternation amongst those of us still defending the last ramparts of human capability, loud criticism arose, given robotics wants to escape the sins of the last three decades:

1. speeded up;

2. selectively successful;

3. videos of performance with no real world relevance;

4. leading to hyped up fundraising; and

5. subsequent trough of despair

(Yes Boston Dynamics, I’m looking at you)

Further examination revealed the Stanford Aloha robot to be a teleoperation framework, ie a human operator teaches the robot using a keyboard how to do the task, then steps back and allows the robot to imitate. The furor caused the lab to release a blooper reel, to soften the FOMO of industry teams who would no doubt be looking to hire the geniuses behind this step up in capability. Thread here.

Do not spill the coffee

While the Aloha bot may have been somewhat miscommunicated in 280 characters, Brett Adcock, former CEO of publicly listed electric Vertical Take Off and Landing eVTOL company Archer Aviation (NYSE:ACHR) and founder of humanoid robotics company Figure certainly did not:

i’m watching robots performing complex tasks entirely with neural nets. AI trained tasks that i didn’t think was feasible until end of the decade

when starting Figure in 2022, i believed we’d have reliable humanoid hardware well in advance of reliable real-world neural nets running on robots

basically, i thought training a robot to do “household” type tasks would size our timeline

my view on this has changed the last several months

i now feel that we’ll likely have reliable AI that can run on hardware around the same time or slightly before the humanoid hardware is highly reliable and starting volume manufacturing

and the path to delivering reliable robotic hardware seems clear and predictable to me, give us time and this will be solved

Timelines are shortening in Robotics as the market end goal: reliable household help, comes into focus. Two days later, announcing Figure had “learned a thing”, Brett posts:

In this video:

• the robot is filmed at normal speed, ie 1x vs the Aloha 6x

• it learned to make coffee by watching a video, ie the loop is visual→motor, rather than the Aloha’s teleoperated motor→motor

• it can correct its mistakes, ie robust to real world differences

Figure must be raising another round right about now, with hiring announcements, videos, etc, and you can’t be seen ceding the memetic race to a bunch of grad students as every VC asks in every pitch “So did you see that Stanford thing?”.

This is how it begins

Our favorite Nvidia data scientist Jim Fan identifies the starting point of the robot takeoff cycle this year, with developments in:

• simulations to provide training data → so that you don’t have to break a billion eggs to learn to make an omelette

• hardware, especially those pesky actuators, sensors and fine motor controls → definitely feel like the 3D printing revolution is helping teams prototype and iterate much faster on customized hardware solutions

• LLMs, especially multi-modal ones, providing human like understanding of the world

• Algorithms bridging the LLMs understanding System-1 to the fine motor control System-2 intelligence

So there you have it, you’re 3 years early to the robot revolution, what are you going to do about that?

Things Happen

• Parental units confiscate the keyboard and mouse? Game on with a tongue activated mousepad. I shall not make the obvious James Bond reference at this time.

• Just in time for the election year, Bland.ai presents an robotic voice caller without a robotic, uh, voice. Five hundred thousand get-out-the-vote calls at the push of a button

• A large focus of post adolescent manosphere inhabitants of the AI space seems to be to figure out a way to reduce the power of the feminine over them, and in true libertarian fashion the chosen method seems to be to alter the supply demand balance. Hence the hunt for the perfect AI model to create the ideal OnlyFans gf continues, with this week’s entries being good enough for a ransom note, or an Instagram page (is that really the same person in all the photos?)

This is the second edition of Self-Aware Neuron! A weekly narrative recap, where I pull together the threads of the week to form a coherent picture from a future historian’s point of view. Subscribe! Tell your friends to subscribe!