Ananye Agarwal

building robot brains @SkildAI | MLD PhD at CMU | Prev CS @ IITD.

[Ananye Agarwal @anag004]

The models we build rarely take us by surprise.

Watching these new capabilities emerge at scale was a rare, unforgettable exception.

[Skild AI @SkildAI]

We built a robot brain that nothing can stop.

Shattered limbs? Jammed motors? If the bot can move, the Brain will move it— even if it’s an entirely new robot body.

Meet the omni-bodied Skild Brain:

[Aradhye Agarwal @AradhyeAgarwal]

Your agent made 15 tool calls. Which ones actually helped?

I built https://t.co/2tUygPDiyd to answer that.

[The Humanoid Hub @TheHumanoidHub]

Skild showed something similar last year but with quadrupeds.

Skild brain, trained on 100,000 diverse simulated robot types enabled remarkable real-time adaptability.

In-context adaptation allows the brain to discern the robot form and adapt to extreme changes in its body.

[Ananye Agarwal @anag004]

Warehouses are a critical part of the modern economy, but are still bottlenecked by human labor.

We've acquired Fetch Robotics to tackle previously hard-to-automate parts of warehouse logistics with our omni-bodied AI!

(Our factorio players are delighted)

[Skild AI @SkildAI]

We have acquired Zebra Technologies’ robotics arm (formerly Fetch Robotics).

This is what happens when orchestration meets intelligence -- a major step toward fully autonomous warehouses.

More robots. More environments. One unified brain.

[Skild AI @SkildAI]

We have acquired Zebra Technologies’ robotics arm (formerly Fetch Robotics).

This is what happens when orchestration meets intelligence -- a major step toward fully autonomous warehouses.

More robots. More environments. One unified brain.

[Ananye Agarwal @anag004]

we're cooking 🧑‍🍳

[Deepak Pathak @pathak2206]

We hosted Prof. Alyosha Efros (UC Berkeley) at @SkildAI! He didn't believe that robots could actually cook eggs reliably. :)
Tested back-to-back 5times without fail! One batch of scrambled eggs every ~2.5mins nonstop. The same model assembles a GPU on a server rack too.

[Deepak Pathak @pathak2206]

We hosted Prof. Alyosha Efros (UC Berkeley) at @SkildAI! He didn't believe that robots could actually cook eggs reliably. :)
Tested back-to-back 5times without fail! One batch of scrambled eggs every ~2.5mins nonstop. The same model assembles a GPU on a server rack too.

[Ananye Agarwal @anag004]

@gauri__gupta @NeoSigmaAI @RitvikKapila congrats!!

[Gauri Gupta @gauri__gupta]

We @neosigmaai @RitvikKapila are building the future of self-improving AI systems! By closing the feedback loop between production data and system improvements, we help teams capture failures, convert them into structured evaluation signals, and use them to drive continuous improvements in agent behavior.

We show how our system works on Tau3 bench across retail, telecom, and airline domains. Agent performance on the validation set (with a fixed underlying model, GPT5.4) improves from 0.56 → 0.78 (~40% jump in accuracy).

[Deepak Pathak @pathak2206]

@yacineMTB Like this?

Our model can control any shape robot, eg. quadruped, humanoid, etc., without being trained on any of them.

[Ananye Agarwal @anag004]

@SkildAI 🚀🚀

[Ananye Agarwal @anag004]

Looking at this task one might ask: why not just motion plan it?

But the fact remains: today, this is done by humans and not robots.

The answer is that real factory lines are messy and static motion plans fail. Robots need to constantly adapt.

The skild brain enables this!

[Skild AI @SkildAI]

Nearly every system today, from energy to chips to food, is bottlenecked by scarce human capital.

We are changing that by building AI-powered industries of the future.

Check out Skild Brain robustly assembling GPU racks, a highly precise task, live at #NvidiaGTC.

[Skild AI @SkildAI]

Nearly every system today, from energy to chips to food, is bottlenecked by scarce human capital.

We are changing that by building AI-powered industries of the future.

Check out Skild Brain robustly assembling GPU racks, a highly precise task, live at #NvidiaGTC.

[Ananye Agarwal @anag004]

The fact that humans can teleoperate existing robots to do a staggering variety of tasks is pretty compelling evidence to the contrary.

Of course, short term progress can (and should) be made by designing hw around limitations of current models. But robotics is a software problem in the sense that, in the limit, generalist models controlling cheap, general-purpose hardware is what will scale best.

[ABB Robotics @ABBRobotics]

Excited to collaborate with @SkildAI to bring AI to industries of the future!

[The Humanoid Hub @TheHumanoidHub]

At GTC 2026 Skild booth, @shikharbahl & @kmarinou_ demo Skild Brain operating autonomously from pixels to robot actions, doing busbar assembly for NVIDIA GB300 compute tray.

Skild uses the same omni-bodied base model for humanoids, quadrupeds, and variety of industrial robots.

[Ananye Agarwal @anag004]

@mattparlmer Hi Matt, this thread has a good discussion on this point https://t.co/YqZGIRlrOy. But TLDR is that hand designed controllers

- scale poorly to lots of tasks since they have to be programmed by experts from scratch
- require low tolerances which are hard to guarantee in general

[Deepak Pathak @pathak2206]

Hey @BradPorter_ @willknight ! Nice to hear from you. Here is what's going on and a few reasons why it is a BIG deal for the future of industrialization.

Unlike classical automation, this is:
- Fully end-to-end neural network (Skild Brain) finetuned with a small amount of robot data (or even zero real-world data)
- Robust to disturbances and extremely fast to set up for any new task
- Long-horizon task with in-context memory. No step-by-step programming
- No custom tuning, no fancy sensors, just vanilla cameras

Why not classical automation:
- Current automation stations require extensive hand engineering and are extremely costly (typically 3x or more than the robot itself). Unfortunately, if the task needs ultra precision, they don't transfer if the setup changes by even 0.1mm, which is very difficult to ensure in a mix of human-operated stations with automation.
- If you look at the live demo for 5-10 minutes, and as we swap racks, you will see the robot needs to adjust the drill or its actions several times due to disturbances in setup. Hence, classical methods fail, so you need a learning approach. Unless you spend a shit ton of money (design for automation DFA, robot in an enclosed cell, heavy sensorization, etc.) to make sure everything is 0.1mm level precise, which beats the point if the setup is going to change quickly. E.g., NVIDIA changes GPU designs every 6months!
- Oh, and this, this is not a made-up task -- it is actually done by humans in Foxconn factory: https://t.co/XTRgr0GLZ2

Vision:
Imagine anyone with no expertise in controls (factory worker, handyman, etc) being able to automate arbitrarily complex factory stations in a couple of days on their own with no fancy setup -- this is going to revolutionize traditional automation!

[Ananye Agarwal @anag004]

We have an exciting update!

We're partnering with the legendary ABB Robotics, Universal Robots, and NVIDIA to massively scale industrial deployments.

Every resource today -- from energy, to chips and even food, is fundamentally bottlenecked by scarce human capital.

Automation is the only path to abundance.

[Skild AI @SkildAI]

Robotics is a data problem.

Today, we’re partnering with @ABBRobotics, @Universal_Robot, and @NVIDIARobotics to deploy the Skild Brain across real-world industries from manufacturing to factory lines.

This will help us build the world’s biggest data flywheel for physical AI.

[Deepak Pathak @pathak2206]

Hey @BradPorter_ @willknight ! Nice to hear from you. Here is what's going on and a few reasons why it is a BIG deal for the future of industrialization.

Unlike classical automation, this is:
- Fully end-to-end neural network (Skild Brain) finetuned with a small amount of robot data (or even zero real-world data)
- Robust to disturbances and extremely fast to set up for any new task
- Long-horizon task with in-context memory. No step-by-step programming
- No custom tuning, no fancy sensors, just vanilla cameras

Why not classical automation:
- Current automation stations require extensive hand engineering and are extremely costly (typically 3x or more than the robot itself). Unfortunately, if the task needs ultra precision, they don't transfer if the setup changes by even 0.1mm, which is very difficult to ensure in a mix of human-operated stations with automation.
- If you look at the live demo for 5-10 minutes, and as we swap racks, you will see the robot needs to adjust the drill or its actions several times due to disturbances in setup. Hence, classical methods fail, so you need a learning approach. Unless you spend a shit ton of money (design for automation DFA, robot in an enclosed cell, heavy sensorization, etc.) to make sure everything is 0.1mm level precise, which beats the point if the setup is going to change quickly. E.g., NVIDIA changes GPU designs every 6months!
- Oh, and this, this is not a made-up task -- it is actually done by humans in Foxconn factory: https://t.co/XTRgr0GLZ2

Vision:
Imagine anyone with no expertise in controls (factory worker, handyman, etc) being able to automate arbitrarily complex factory stations in a couple of days on their own with no fancy setup -- this is going to revolutionize traditional automation!

[Lukas Ziegler @lukas_m_ziegler]

Ok so @SkildAI just taught a robot an industrial task in 10-15 hours

That’s awesome! GTC keeps surprising ;)

[Chris Paxton @chris_j_paxton]

This is the first time that I have seen Skild doing live demos in public! Demonstrating neural nets for precision manufacturing