This page will serve as my general notepad for all things Robot, AI Robotics, and “embodied AI” (which appears to be the latest buzzword to express the idea… an AI with a physical body, we get it). So I will be updating it regularly, for those interested.
How AI Robots learn by trial and error
This (video, below) demonstrates elegantly (and primitively!) how modern robots learn — just like their ChatBot brothers — they are not “programmed” or “coded with rules” or given infinite “if/then” conditions… they are simply hyper-accurately modelled, down to the screw, joint, mass and spring constant, then those models are thrown into a simulated training ground with hyper-accurate physics modelling (gravity, terrain, obstacles, objects, etc), and given 40 million minutes of training to… learn.
That is, to figure it out for themselves. Importantly, there is a goal that they are oriented towards, and a success metric that they are fed in order to judge their success (or failure)… in this case, the goals are something on the order of:
- achieve a high degree of mobility,
- attack with your weapon,
- defend with your shield,
- be the last ‘bot standing
- (i.e. defend yourself, vanquish your enemies, stay alive).
Learning in Sim: from crawl to combat
It is utterly fascinating to watch this first simulation, where 100 robots literally faceplant, then watching as they evolve into crawling, walking, running, and finally (after the aforementioned 40 million trial-and-error runs) evolve into some acrobatic combat maneuvers that would put any self-respecting ninja or stuntman to shame.
Observe:
Think of it just as young drivers are taught in car simulators, and pilots are trained in flight simulators, long before they are allowed to fly a real plane with passengers (or missiles).
The simulations are so physically (read: physics / gravity) accurate, that the learning can be directly translated into “embodied AI” systems… that is to say, real world humanoid robots.
Boston Dynamics / Atlas: the Military’s AI Robot
This is precisely the work that a famed company called Boston Dynamics does. This is the result of all that AI simulation learning, embodied in their most advanced humanoid robot model, Atlas:
Atlas doesn’t just play in controlled warehouses. It is, from time to time, let loose in the wild to test its autonomous navigation and power duration:
UPDATE: Feb 2023
Natural Language control of robots
in non-standard environments.
Atlas is, then, the real world application of all that training in simulation
To see where it all began,
from what humble origins,
here’s a home video of
our baby AI’s very first steps,
oh so long ago, way back in 2017:
TeslaBot : an AI Robot with an MSRP?
update, Jan 2023:
Some thoughts on the TeslaBot:
of all people, it was Tony Stark himself who prompted me to re-evaluate this strange pivot into AI Robotics from an already very successful (understatement!) enterprise:
and then,
from Tesla’s
Optimus v2.0 reveal:
“If you think about it,
we’re just moving
from a robot on wheels
to a robot on legs.”
— Milan Kovac
Project Lead, Tesla “Autonomy” Team
The Top 4 Humanoid Robot Initiatives as of 2023
A quick nod here to all the worthy predecessors (and truly, present day siblings) of the teslaBot:
The Use Case for Humanoid AI Robotics Platforms
“[The first generation of humanoid] Robots, in the short term, have use cases in places where they do not necessarily need to interact with humans and the implications of failure are less severe.
The community needs to find a revenue-positive pathway to support this development. And this could come from behind-the-scenes use cases for robot manipulation, in warehouses, retail stores, food preparation, and manufacturing. While automation-based solutions are still being pursued, it remains to be seen how a general-purpose hardware-based solution (i.e. a task-adaptable, humanoid-form-factor robotic crew) would stack up. We would need to look to low-volume, high-variability products which require quick adaptation. (as opposed to high-volume, low-variability products, like car manufacturing, which favor more task-optimized large-scale “industrial” class robots).”
— Animesh Garg
What Roboticist Experts think of the TeslaBot
AI Robotics : The Secret Ingredient
“I’ll let you all in on a secret about humanoid robots:
It’s all about reliability.
How often does it fall down?
You can’t tell from a cool video—or even a live demo…”
— Christian Hubicki
ibid
G: …You can only tell from thousands — actually, hundreds of thousands — of hours in the field, in actual, real-world application. Then we’ll know. Until then, we fantasize, and dream.
But don’t worry, your robot on four wheels (your camera-adorned, cloud-connected automobile which you think you own) is an embodied AI robotics system, and is collecting that data 24/7, and is using its 5G-cloud-uplink to stream all that data to the AI Overmind, where it is being carefully analysed, appended, and added to the global knowledgebase… not only of “how to drive a car” but also of “what the world looks like” and perhaps most importantly “how people move through the world”
And finally, we depart this post on a powerful visual note:
A Vision of a Robot Factory, 2020
Nope.
They certainly ain’t humanoid. (giraffoid, perhaps?)
But they’re definitely robots.
Next up:
Why the AI doesn’t need robots… yet.
…it has millions of humans to perform its every whim.