Darius Foodeei

ML @ETH Zurich prev. @EPFL | @ethroboticsclub Humanoid software lead

[Darius Foodeei @dariusfdi]

@DominiqueCAPaul seems like it's not necessarily just a question of safety regarding humans it's also a question of enforcing guarantees that make deployments predictable

[Darius Foodeei @dariusfdi]

Do all these robot learning companies have certification and safety plans?

[Darius Foodeei @dariusfdi]

@dimensionalos @stash_pomichter @ruthwikdasyam very cool! Have you guys looked into using kyber?

[Darius Foodeei @dariusfdi]

@awscloud please. more capacity for g7e instances

[Darius Foodeei @dariusfdi]

gotta hand it to @OpenAI 5.6 sol is excellent

[Darius Foodeei @dariusfdi]

@DominiqueCAPaul really? afaik needs to be enforced between the policy and the controller

[Darius Foodeei @dariusfdi]

@TheHumanoidHub iron man inspired me

[Darius Foodeei @dariusfdi]

@DominiqueCAPaul current state of robot learning rn lol. but 2027 is coming soon and the EU Machinery Regulation makes it illegal to sell a machine with a learned policy on it in Europe without a conformity path

[Darius Foodeei @dariusfdi]

@ErenChenAI 4) What?

[Darius Foodeei @dariusfdi]

@OmeadMoses It is and it’s highly underrated

[Darius Foodeei @dariusfdi]

This is very true. Every lab and major robotics company I've spoken to is trying RL orchestration via LLM / VLMs, VLAs (usually pi0.5 / Groot style) and world models (most of the time just Cosmos + IDM no JEPA) all at the same time. no one really seems to be set on one approach

[Ethan Clark @ethanmclark1]

Working in robotics right now is what I imagine working with language models felt like in 2023. Everyone throwing things at the wall to see what sticks

Pixel prediction (Cosmos), action prediction (VLA), reward prediction (TD-MPC), and representation prediction (JEPA). Different paths for the same problem

The recipe that won in language was self-supervised pretraining at internet scale then light finetune on top. Only representation prediction runs that playbook. It learns from action-free video data so you can pretrain on YouTube and egocentric data then add a control layer. Everything else needs action-labeled data that doesn't scale

As an RL maximalist, I used to hate LeCun's cake. Turns out he was right all along which is how I ended up a JEPA truther

[Darius Foodeei @dariusfdi]

@radbackwards Do those have tactile sensors?

[Darius Foodeei @dariusfdi]

For such a bare metal build I have to say the cabling work is very clean. Cables when daisy chaining dozens of actuators is a real and highly underrated challenge. Impressive!

[Remi Cadene @RemiCadene]

Starting with the fundamentals

Prototype Version 0
AI, Software, Hardware
A small team, 9 months
Designed and assembled in Paris at @UMA_Robots

[Darius Foodeei @dariusfdi]

@RemiCadene @UMA_Robots Very much like the sheet metal body. We’ve taken a similar approach @ethroboticsclub for our first humanoid

[Darius Foodeei @dariusfdi]

@baraqlipshitz Had exactly the same issue with the G1

[Baraq Lipshitz @baraqlipshitz]

Project Update: I’m training a World Action Model for general bimanual manipulation.
Here’s everything I’ve learned so far, including the CPU contention bug that was silently corrupting my data. Thanks to @nvidia, @AWS for the compute and @ethroboticsclub for the hardware! (1/8)

[Darius Foodeei @dariusfdi]

Will be at @MACHINASUMMIT tomorrow happy to meet anyone working on robotics and world models.

[Darius Foodeei @dariusfdi]

@baraqlipshitz Usa and usa

[Darius Foodeei @dariusfdi]

@DominiqueCAPaul Hopefully I’m not too late!

[Darius Foodeei @dariusfdi]

Curious to see what @Figure_robot @1x_tech @Tesla_Optimus and others can show off in terms of long horizon tasks like this 👀 Very impressive and all out of Zurich 🇨🇭

[Flexion @FlexionAI]

"A parcel with snacks has been delivered for Flexion. Retrieve it using the stairs and come up using the elevator. Then unpack it and place the items into the empty drawer on the shelf in the snack area."

One instruction. No human operator. Everything that follows is autonomous.

Today we're introducing Reflect v1.0, our robotics intelligence platform for long-horizon work. From a single natural-language command, the robot understands the task, navigates a multi-floor building, calls elevators, handles doors, uses tools to unpack a box, and puts the items away. The biggest shift in v1.0 is that we use reinforcement learning across every layer, from low-level control to high-level reasoning.

Long-horizon autonomy is unforgiving. The robot must recover on its own when things don't go to plan because in the real world, they never do. Combining reasoning, perception, physical execution and runtime robustness into a single mission-capable system is the foundation required to solve humanoid autonomy.

Our team is just getting started.

#HumanoidRobots #Flexion