This is the recipe to build robotics at scale. Own the end customer relationships, build the data engine, scale robots for your own margin improvements. It is our approach at
Potentially nitpicky but competitive advantage in AI goes not so much to those with data but those with a data engine: iterated data aquisition, re-training, evaluation, deployment, telemetry. And whoever can spin it fastest. Slide from Tesla to ~illustrate but concept is general
Robotics folks - how on earth does a single celled creature know how to morph what is essentially a part of its body into an arm, and grab a tiny object in 3D space? What localisation, mapping, planning and control is it running (if any - or is it purely random)? What makes this
An efficient and vicious microscopic hunter, the single-cell organism Lacrymaria olor, attacks and consumes another single-cell organism.
📽: James Weiss
Great summary on scaling laws wrt robotics.
However - i would add that the robotics tech community would seriously benefit from a discussion on business models that can profitably explore this question.
Beyond the dichotomy of arm farms/ (insert your
The hard part about robotics and autonomy - humanoids, robodogs, AMRs, AVs isn't hardware. The hard part is to build a data engine to achieve agency that is on par with human reliability - and to do it profitably. This is what
@elonmusk
and
@Tesla_AI
and have built.
Step 1 is to
An alternate vision of the future is one of "robot animals" - not the FUD of terminator AGI. As an example - consider this video of two
@sheeprobotics
robots servicing a park in CA, as cobots alongside a human.
What's interesting here is that there is no setup or teaching -
Robots have to earn their right to exist in the real world. What is preventing end2end ML robotics from becoming reality is a lack of demand beyond one-off pilots.
This is driven by a lack of usefulness,
-> driven by a lack of robustness
-> driven by lack of
Super stoked to see Electric Sheep
@sheeprobotics
mentioned in the
#GTC2024
@nvidia
GTC keynote today! Come visit our booth tomorrow and day after at GTC!
Here's an application for our ES-1 world model that i'm really passionate about - targeted weed treatment.
Today, outdoor service companies takes a "spray and pray" approach to most kind of outdoor chemical application - this results in an indiscriminate use of harmful
EXCLUSIVE: CEO Unveils AI Machines Beyond Tesla's Reach! | Electric Sheep Nag Murty
Can LLM-Style robots revolutionize the physical world?
@sheeprobotics
My car drives itself.
My computer talks to me.
My garden is abuzz with robot helpers.
Every now and then, when I am not consumed by my own journey, i stop, and am amazed that this should all even be possible ...
Unscripted reasoning and planning leading to Unsupervised interactions with Unstructured worlds is the Ultimate AI problem to solve.
Robotics engineers have been climbing this hill for years now - all the SaaS bros will now catch up to the harsh realities of making an "agent"
ChatGPT + robotics.
How ChatGPT struggles with the same things we've been struggling with in robotics.
With
@MurtyNag
and Mike Laskey from
@sheeprobotics
.
Had a lot of fun on this with
@audrow
, and we're delighted to be his first guests on his new podcast! If you're into all things robotics - Audrow is a must follow!
Does anyone else see the parallels between these AI agents like rabbit R1, humane pin, coding agents like Devin and embodied AI robots?
They all have an "interaction data" bootstrapping problem - and unless they build a product which delivers value independent of AI (Tesla EV vs
when do you think we'll have robot arms that can do this?
Today's tech seems positively medieval compared to what biology has built up with a fusion of sensors and actuators at the cellular level.
And until we do - can robots ever truly be as efficient, versatile and
the hard part about ML and robotics isn't the flashy demo - its the grind and building a real business that can extract value from an imperfect product that gets useful over time.
this is incredible - how do other humanoid companies even hope to beat this combo of manufacturing, internal use within their own factories, and the foundational ML models they've gained from the Tesla fleet?
RaaS is a dead end and a money pit (for the foreseeable future).
arms make for cool demos - but the simpler the hardware - the better it is to solve for the data flywheel and robust end to end ML robots.
we should seek to solve for navigation before we solve for manipulation.
As a robotics community - we are severely underestimating the
@Scobleizer
@sheeprobotics
It was great to have to you over!
@Scobleizer
Boring businesses are indeed great for Embodied AI - they allow us to build a real world sandbox, make money and chase increasingly complex robotics over time!
100% agree. Imagine a swarm of robots that allow you to "edit the earth" around you on demand ...
outdoor robotics is a photoshop-tool to express human creativity on planetary scales.
Surely there isn't a startup (1) acquiring landscaping businesses, (2) to build a data engine, (3) to train embedded AI models, (4) so they can develop novel robots...
Episode 33 of S³ features
@sheeprobotics
Sharing a bit more about Reflect Orbital today.
@4TristanS
and I are developing a constellation of revolutionary satellites to sell sunlight to thousands of solar farms after dark.
We think sunlight is the new oil and space is ready to support energy infrastructure. This
chatGPT is nice … but we are so ridiculously far from building catGPT - animal-like nonverbal intelligence that can reason and plan in 3d space …
To quote
@ylecun
A house cat has way more common sense and understanding of the world than any LLM!
Who would you rather be: the robotics startup that hopes some incumbent will select you after pilot purgatory, or a behemoth that rolls up and dominates an entire industry?
here’s popular mechanics in. 1949:
Where a calculator on the ENIAC is equipped with 18,000 vacuum tubes and weighs 30 tons, computers in the future may have only 1,000 vacuum tubes and perhaps weigh 1.5 tons.
@Suhail
great point. Also as agentic AI becomes the end goal, owning the data flywheel is the only moat that matters. There is very limited ways (outside of simulated data) for these models to get access to perception-decision-action feedback loops that can only happen as you deploy
We trained ANYmal to go into confined spaces such as under collapsed buildings. To be presented at
#ICRA2024
Title: Learning to walk in confined spaces using 3D representation
Arxiv:
Video:
Summary Page:
@audrow
the key challenge is that going from arduino/Pi to a reliable robot is a huge leap. It's easy enough to build some basic hobby robots. But the leap from that to production seems to need insanely disciplined engineering, and very little technical brilliance (relatively). This
Watch as two Electric Sheep robots service a park while the other crew member takes care of the edges, and other dextrous tasks like blowing, and trimming!
Here's the power of AI at work in the real world - a human + robot crew that can take care of some really complex turf
nature is incredible. building animal-like AI robots will also mean alignment - how do we realize a machine which has a drive to self-preserve, but also empathy?
@audrow
tangential - but AR glasses paired with fine tuning multimodal models can pull a lot of people into robotics to generate task POV datasets
without having to get into grimy hardware 😊
@natfriedman
we're not going to get to AGI without embodiment and grounding in real world physics. This will require a means to deploy robots at massive scales and create Tesla-like data engines.
Someday, we'll drive down the cost of manufacturing to zero with robots and nano factories. And every human can manifest such a personal reality - especially those caught up in the soul-sucking grind of subsistence living in large parts of the world.
So this short article has been doing the rounds on twitter lately ... If I'm reading it correctly - it implies data quality is the only thing matters - all other investment tends to commodity in the limit (GPUs, people, architectures etc) - what implications does this have for a
This will change when the VCs realize that services companies shouldn't just valued for cash flows (a very traditional PE worldview) - but for the "compressed intelligence" value of the petabytes of interaction data that they generate everyday.
im surprised how few VCs seem to be aware of the “took an operationally intensive business and made it high margin and scaleable with llms” plays
afaict they’re the largest category of llm uses that make money or have any traction
@AbhayVenkatesh1
At We're building the openAI equivalent for the physical world. But our key insight is that to do this - we will need to embrace a vertical integration play where we buy services companies to own the data, and the means to do RL under the air-cover of a
This is an interesting question to ponder. Will robots be superhuman or human-like?
Another lens to maybe view this through is this:
Is the task already mechanized, or do human limbs provide much of the motive force.
Where mechanized - no point in a humanoid (e.g. Zero-turn
Should humanoid robots drive manual cars or should we have self-driving cars?
Should humanoid robots push our manual lawn mowers or should we have a robotic lawn mower?
Should humanoid robots scrub dishes with a sponge or should we have a dishwasher?
Should humanoid robots
@hausman_k
not really - robotics is best viewed not as a product but as a learning system. One shot investments wont fix the underlying challenge which is:
1. using an 80% good-enough robot to deliver value so that ...
2. enough data can be generated to solve the long tail that is the
@BorisMPower
animal intelligence co-evolved with the body each one incrementally shaping the other. In robotics, we're actuator and embodied data constrained - both will need massive improvements to crack Moravec's paradox.
real world data from real world feedback - not just passive observation.
its easy to simulate or passively collect data.
hard to collect feedback from millions of early adopter users who use your product enough to subsidize your innovation.
Some thoughts on the pillars of full self driving:
1. The inference chip inside the car. Who has it? Nvidia's Drive Thor, MobilEye EyeQ, Tesla FSD, Waymo Drive Compute, Nio Shenji
2. The training compute needed. Who has it? Nvidia's H100 / A100 that everyone relies upon.
@TashaARK
@skorusARK
terrific! this is a template for much of robotics - owning data and distribution and a really large RL sandbox before one can start to license features or sell robots.
ML systems need time to bake to SLA levels of performance. Current business models in robotics don't allow for
Language Models dont have a data problem (web scraped + synthetic), and aren't edge constrained at inference - it makes sense they are going to max out on compute.
Robotics on the other hand - is heavily data constrained AND edge constrained. As an industry - we should be
imo. totally worth it. We spend 100x more as a society on trivial incremental innovations.
solving for unlimited intelligence, energy and labor - can end human suffering and poverty.
OpenAI's Sam Altman: Whether we burn $500 million a year, or $5 billion or $50 billion a year, I don’t care. We’re making AGI, it’s going to be expensive. It’s totally worth it.
$50 Billion per year for AGI. Is it really worth it???
@audrow
LLM agents = software robots. Anyone working on agents will soon come to the conclusion that data and RL across all steps in a given task will be key.
That data is missing today from the web-scraped info available. It's only to be found locked away inside boring companies
A video of this robot has taken Instagram by storm! 🌪️
This video shows a robot that was asked to serve a can of Coke.
The robot sets about completing the task but dropped the can on his first attempt. After realizing that he had failed to complete the task, he decided to
The earliest attempts to build a car didn't look anything like a car.
Inventors in the 1860s were convinced that the next leap in mobility would be obtained by building a steam-powered anthropomorphic walking machine that pulled a carriage. And so the steam man was born.
This is a very relevant way to think about robotics data as well -
It's not about passive data collection, but a data feedback loop with deployed models and users that experience them - creating a data engine for ongoing improvement
It was instantly overshadowed after the stock went hyperbolic, but during the earnings call, Mark Zuckerberg mentioned that the gargantuan trove of public data they have amassed from Facebook and Instagram to train Llama 3 exceeds the Common Crawl. So, in excess of 6.4 petabytes.
@BrianJJi
This is also playing out with AI. Most companies are looking at AI as a product to make revenue by selling to incumbents - pattern matching off of SaaS.
The 100X bigger opportunity is to take boring old industries and consolidate the smaller players within it and use home-brewed
Rather than thinking of AI agents in terms of some omniscient AGI, we need to look past the FUD and marketing hype and see them for what they are: assistants who will learn alongside their human teachers.
@tszzl
the biggest bottleneck to this adoption is behavior change and organizational inertia - especially if the AI is more cobot than fully autonomous agent
wow - this is clever. stay within simulation, and generate content based on AI agent interactions within the simulation.
Eventually - This is the way to photorealistic content - just swap cartoony southpark for unreal engine.
Announcing our paper on Generative TV & Showrunner Agents!
Create episodes of TV shows with a prompt - SHOW-1 will write, animate, direct, voice, edit for you.
We used South Park FOR RESEARCH ONLY - we won't be releasing ability to make your own South Park episodes -not our IP!
@Plinz
Don't Buddhism and Advaita both maintain that everything animate and inanimate are just the means for the universe to experience itself? The cosmic ocean will roll on in its creation and destruction - regardless of what happens to moby dick or ahab or ishmael. Neither Individual
@audrow
i think the key is real world RL. The doggie on the left turns into that beast on the right after many, many feedback loops in the real world, at scale across hundreds/ thousands (?) of robots. Robotics is all data engine, less about the first sim-2-real prototype. That's just a
@DrJimFan
If we have to climb out of paper purgatory and demo dreamscaping - we have to address the big elephant in the room - which is that ML that will drive robots needs to be really precise, fast, safe and have positive ROI to the operations deploying it.
We need to be solving for
Multimodal AI agents are the only thing worth solving for. Hardware or Software is just a means to actuation.
Agents make obsolete everything else in the limit - all SaaS (human operated software tools), all machinery (human operated hardware tools)
we should all just skip
This video is misleading - the horse has enough common sense to distinguish between painted vs real grass - and they use multi-modal inputs to achieve this - vision being only one of them.
Animal intelligence is really robust to most variations in the environment, and is a great
Horses are not color blind, they have two-color, or dichromatic vision, but have relatively poor "accommodation" (change focus), as they have weak ciliary muscles
Anyway, this painted grass was enough to fool this horse.
agree 💯building and backing audacious goals is what makes Silicon Valley special. Solving the agency problem is very hard - all paths to figuring out a viable data flywheel and business model ought to be cheered and explored - healthy critique will help the community
I'm seeing the word "grift" being thrown around too much lately.
I have a Rabbit. It's a real product in my hand.
Now whether it is good or not is a totally different subject, but the term "grift" should be saved for people who take your money and don't deliver anything in
Wen catGPT? this has to be the most visual example of moravec's paradox. Robotics needs a breakthrough in "physical intelligence" - both in terms of hardware and software.
@audrow
Thanks for the shout out! Really appreciate it!
@sheeprobotics
We'd love to share our thinking and approach for others who are considering this in different industries.
so excited to see the Embrace incubator towards the end
@vkhosla
- honored to see you include it in the video.
I have every single atom of that product etched into my brain - lived breathed and fussed over every minor detail!
@MPrinParr
@jasonjoyride
@sheeprobotics
@audrow
Very thoughtful feedback and insights here - While we can get very far with the current Nvidia architecture - the real unlock for robotics will come from very energy efficient neuromorphic compute like
@Rain__AI
@gordonhwilson
. Animal brains and bodies are the north star for
@simonkalouche
Until you can fuse the sensor, actuator, energy sources into each other at the lowest level of construction like biology has - bionic robots will be highly suboptimal and a marketing tool at best.
When Spencer Seabrooke broke the world record for the longest free solo slackline walk.
He used no safety gear at all, walking over 200 feet across a 1,000-foot high canyon in Canada.
@chris_j_paxton
@sheeprobotics
Thank you for the shoutout! As ML eats more of the robotics stack - building a learning system becomes the key focus. Especially if the goal is to build a world model that can generalize to all outdoor tasks.
Assuming bipedal = humanoid: This needs a few things to happen -
1. Some key advances in embodied ML (a new paradigm beyond LLMs)
2. A new kind of (energy efficient) compute beyond GPUs
3. A new means to actuate beyond the current servo/ cable approaches - that is much
The thing nobody talks about is that in 10 years we'll have a million bipedal robots and in 25 years we'll have a billion. You’ll buy yours for $10k and it will be as important to your life as your smartphone is now
We're basically going to have to "parent" an AGI similar to how we raise our children and train our apprentices: collectively as a society.
I thought
@sama
's
@theallinpod
interview was interesting and pointed towards this future.
This implies:
0. We're going to spend the next
@ericjang11
Great read! Would've been great to see some discussion around data engine economics - building a profit-generating flywheel seems to have been the only approach that has worked for prior successes (Amazon and Tesla). The economics of the data engine should drive the embodiment
@audrow
we’re always automating some task with every new piece of technology we’ve invented- since the stone axe.
but the need for new jobs/ people seems to grow every time - because productivity gains unlock more elaborate and grander human needs - the size of the pie is the whole
2023 was a huge year for progress in robotics and manufacturing. I cover it all, including:
- US manufacturing boom
- El Segundo, the new Silicon Valley?
- Humanoid, AVs
- China EVs dominate, especially in Europe
- market consolidation and mega-rounds
(link below)
@audrow
Another interesting point here is - if they build robot for X USD -> that just translates to an X USD cost to them - because its an internal tool.
If they bought from one of the numerous AMR companies - it would likely be 3X price (because these companies have to justify their
@DrJimFan
to build robotGPT - we need a business model that can drive data collection, deployment and RL at the scale of hundreds of thousands of deployed robots.
tesla has managed to do this for AV - and at Electric Sheep, we’re doing this for outdoor robotics.
@oyhsu
We think it runs through any business that can create a data flywheel for an end to end model to mature and gradually provide business value.
General robotics is a misnomer - for a while we are going to be tied to domains (outdoors, indoors, warehouses etc) and embodiments
The petri dish of reality + billions of years of evolution gave us this ...
To build robots that can come anywhere close to this level of adaptation and perfection - we will need to dream up new petri dishes in simulation and business models to fine tune in reality.
both e/acc and decel are incredibly stupid megaphones with which to opine on the future (or explain the past).. Individuals have as much choice in acceleration or deceleration of technology as a tea leaf had in the evolution of the east india company.
Every solar panel is like a giant pixel in a camera sensor, turning photons into a voltage. With billions of them across the world, it is interesting to imagine what they could collectively “see” of the sun and atmosphere, even though they have no focusing machinery.
Conversely,
@bindureddy
Word. Automation will imply agency (reasoning and planning vs retrieval) - this is a tough nut to crack, and arguably the only thing worth solving - since it renders everything else obsolete!
@MPrinParr
humanoids will need to bake within a real use case and be perfected. Since Tesla plans to use them within their own factories to start with - they will be best positioned to make them robust.
Most Robotics as Service companies today directly or indirectly subsidize deployments
@justindross
Sounds a bit similar to running a landscaping business. I'm running a robotics company that's executing a PE style rollup of landscaping companies (for data, RL and distribution). Have developed a lot of respect for what service industry entrepreneurs have built over the decades