@IlirAliu_: Forget task-specific robots. ...
@IlirAliu_
11 views
Oct 05, 2025
1
Forget task-specific robots.
We're entering the era of Robotics Foundation Models (RFMs), AI brains that will unlock truly intelligent robots.
Why it’s a game-changer…
and the players behind it 👇
[A video deep dive. Save it for later 📌]
We're entering the era of Robotics Foundation Models (RFMs), AI brains that will unlock truly intelligent robots.
Why it’s a game-changer…
and the players behind it 👇
[A video deep dive. Save it for later 📌]
2
For decades, robotics was... dead. Every robot was a custom-programmed puppet, only capable of doing one thing.
A factory arm couldn't cook dinner. A delivery drone couldn't walk a dog. They were siloed, dumb machines.
[Still, that KUKA’s pretty cool.]
A factory arm couldn't cook dinner. A delivery drone couldn't walk a dog. They were siloed, dumb machines.
[Still, that KUKA’s pretty cool.]
3
The pioneers building RFMs.
Imagine an AI brain so smart, it can power ANY robot
(humanoids, quadrupeds, drones) and teach them to do anything with minimal training.
This isn't incremental. It's a paradigm shift.
[Clip from @SkildAI; more companies below 👇]
Imagine an AI brain so smart, it can power ANY robot
(humanoids, quadrupeds, drones) and teach them to do anything with minimal training.
This isn't incremental. It's a paradigm shift.
[Clip from @SkildAI; more companies below 👇]
4
Robots are learning like humans now.
Not from code, but from experience.
Millions of interactions. Human demos. Billions of simulations. That’s how they gain common sense.
[RT-X, the largest open source robot dataset ever]
Not from code, but from experience.
Millions of interactions. Human demos. Billions of simulations. That’s how they gain common sense.
[RT-X, the largest open source robot dataset ever]
5
Leading the charge:
@physical_int (π0): The "OpenAI for robotics," building generalist brains and open-sourcing their tech.
@SkildAI (Skild Brain): Training on data 1,000x larger than rivals, teaching robots everything from climbing stairs to assembling objects.
@physical_int (π0): The "OpenAI for robotics," building generalist brains and open-sourcing their tech.
@SkildAI (Skild Brain): Training on data 1,000x larger than rivals, teaching robots everything from climbing stairs to assembling objects.
6
Then there's:
@nvidia (Isaac GR00T N1): An open-source brain for humanoid robots, with a dual-system AI architecture like the human brain.
@CovariantAI (RFM-1): Powering warehouse automation, learning from billions of real-world logistics interactions. (Amazon already bought their tech AND team! 🪦)
@nvidia (Isaac GR00T N1): An open-source brain for humanoid robots, with a dual-system AI architecture like the human brain.
@CovariantAI (RFM-1): Powering warehouse automation, learning from billions of real-world logistics interactions. (Amazon already bought their tech AND team! 🪦)
7
While some build for the lab, @DynaRobotics builds for the real world.
They just raised $120M for their DYNA-1 model, which already boasts a 99% success rate in 24/7 continuous operation.
[Here is a @corl_conf 2025 timelapse]
They just raised $120M for their DYNA-1 model, which already boasts a 99% success rate in 24/7 continuous operation.
[Here is a @corl_conf 2025 timelapse]
8
But what about the millions of robots already out there? @mbodiai has a different approach.
They're not selling new hardware. They're building the AI layer that integrates with existing robots, giving them generative AI brains and natural language skills. It's an upgrade, not a replacement.
[More from them in a couple of weeks 🤓]
They're not selling new hardware. They're building the AI layer that integrates with existing robots, giving them generative AI brains and natural language skills. It's an upgrade, not a replacement.
[More from them in a couple of weeks 🤓]
9
And then there are the purists. @genrobotics_ai is tackling the grand challenge head-on: building true general-purpose intelligence that works across any form factor.
Their mission isn't just to automate tasks, but to unlock a universal physical intelligence.
(next week w/ co-founder @akapoor_av8r on my pod)
Their mission isn't just to automate tasks, but to unlock a universal physical intelligence.
(next week w/ co-founder @akapoor_av8r on my pod)
10
Why now?
Data: robots finally generate enough of it.
AI: LLMs showed how to build generalist models.
Sim: tools like Isaac Sim scale training infinitely.
A perfect storm. ⛈️
Data: robots finally generate enough of it.
AI: LLMs showed how to build generalist models.
Sim: tools like Isaac Sim scale training infinitely.
A perfect storm. ⛈️
11
This isn't just about making robots smarter. It's about making them adaptable.
Instead of reprogramming for every new task, you tell an RFM-powered robot what to do in natural language, and it figures it out.
Laundry? No problem. Assemble a box? Done.
[Intelligent picks and perfect fits; Covariant]
Instead of reprogramming for every new task, you tell an RFM-powered robot what to do in natural language, and it figures it out.
Laundry? No problem. Assemble a box? Done.
[Intelligent picks and perfect fits; Covariant]
12
The biggest bottleneck in robot AI? Data.
Real-world data is scarce and costly... so companies are getting creative:
• Fleets that learn together (Covariant)
• Massive virtual worlds (Skild AI, Agility)
• Cross-embodiment brains (Physical Intelligence)
[whole-body controller trained in Isaac Sim w/ RL]
Real-world data is scarce and costly... so companies are getting creative:
• Fleets that learn together (Covariant)
• Massive virtual worlds (Skild AI, Agility)
• Cross-embodiment brains (Physical Intelligence)
[whole-body controller trained in Isaac Sim w/ RL]
13
The new robotics giants are pushing:
@BostonDynamics & Toyota: Developing Large Behavior Models for Atlas.
@Figure_robot: Building Helix, their VLA model, after raising $1B+ from Bezos & Microsoft.
@agilityrobotics: Training whole-body control models for Digit in simulation.
@BostonDynamics & Toyota: Developing Large Behavior Models for Atlas.
@Figure_robot: Building Helix, their VLA model, after raising $1B+ from Bezos & Microsoft.
@agilityrobotics: Training whole-body control models for Digit in simulation.
14
Over $7B went into robotics foundation models in 2024.
A bet on physical AI.
The winners won’t have the flashiest robots.
They’ll build the smartest brains.
The body’s the vessel. The brain’s everything.
A bet on physical AI.
The winners won’t have the flashiest robots.
They’ll build the smartest brains.
The body’s the vessel. The brain’s everything.
15
Robots won’t just execute tasks... they’ll understand and reason about the world around them.
From factories to homes, this shift will redefine how we live and work.
Are you ready for robots that learn and adapt like humans? Which RFM company will lead the way?
[@foundation_robo]
From factories to homes, this shift will redefine how we live and work.
Are you ready for robots that learn and adapt like humans? Which RFM company will lead the way?
[@foundation_robo]
16
If you found this interesting, repost the tweet below, and follow me @IlirAliu_, a thread like this every Sunday!
View Tweet