Vad | Exylos Robotics

Turning human intent into validated robot-ready manipulation skills for complex high-variability tasks. @exylos_ai

[Vad | Exylos Robotics @vad_shev]

@andreasklinger Centralized data farms just train robots to be really good at working in centralized data farms. The real world has too much entropy for this to scale.

[The Humanoid Hub @TheHumanoidHub]

Sam Altman on the Sources Podcast: OpenAI 'will definitely do humanoids.'

It lines up with OpenAI's actual moves:

• It's been building an in-house robotics team, with 33 roles currently open in SF, spanning actuator design, data-acquisition ops and embodied AI.

• OpenAI shut its original robotics team around 2021, citing a lack of real-world data to train useful systems. Now they're going all in.

• Sam Altman has repeatedly signaled his leaning toward the humanoid form in past interviews.

OpenAI raised $122B in its latest round, so it has the deepest pockets in the field and pays top dollar for engineers. It's the least talked-about humanoid player, largely due to its secrecy, but likely to be a major force once it fully emerges.

[Vad | Exylos Robotics @vad_shev]

@lukas_m_ziegler @Gobanorobotics Funny, we get the same thing at the data stage. Clip by clip everything passes, and then the batch turns out to be the same task recorded over and over.

[Vad | Exylos Robotics @vad_shev]

@XSquareRobot 5.3x lines up with what we see. What got harder for us at volume was the people. Fifty contributors on the same setup, recordings that barely resemble each other.

[Vad | Exylos Robotics @vad_shev]

@YunzhuLiYZ @jcjohnss @BenMildenhall @drfeifei The few recordings to controllable scene part is what I'd want to try. Interesting what happens once a robot acts inside it, whether contacts and object motion hold up when a policy pushes and grasps things.

[Vad | Exylos Robotics @vad_shev]

Another simulation sample dataset is on @huggingface 🤗
Proud of our team!

[Exylos @exylos_ai]

Sorting objects is one of the most requested skills from robotics teams — and far harder than it looks.

Moving a thing from A to B sounds trivial.
The hard part is the range: a huge variety of objects the robot has never seen, telling them apart, gripping each shape, handling the mess of a real cluttered scene.

Covering that range with real-world capture is slow and expensive. Simulation generates the variety faster and cheaper — enough to build a dataset that's actually useful.

New sample's public on Hugging Face now.
Link in the comments.

[Vad | Exylos Robotics @vad_shev]

@lukas_m_ziegler Parity solves the geometry and locks it at the same time. Every hardware iteration means the demos you already recorded describe a slightly different robot. Curious how RAI treats the older data.

[Vad | Exylos Robotics @vad_shev]

@sincethestudy If you're ahead on this, is it because you condition on something beyond pixels, or because your data already carries the contact information?

[Vad | Exylos Robotics @vad_shev]

@SongShuran Reads like the buffer's composition matters more than its size. If that's right, the practical question for anyone collecting data is which fraction to keep, not how many hours to record.

[Vad | Exylos Robotics @vad_shev]

@Lindon_Gao What drove most of the Dyna-1 to Dyna-2 jump, field failures going back into training or changes on the model side?

[Vad | Exylos Robotics @vad_shev]

@antopatrex1 Deployment hours are a strange unit to measure a lead in. 10,000 hours of one fleet doing one task in one kind of site is a narrow slice, and everything outside it stays invisible either way. Real head start, but the coverage gap is much smaller than the hours gap.

[Vad | Exylos Robotics @vad_shev]

@aliattar This is what a dream job looks like 😌

[Vad | Exylos Robotics @vad_shev]

@aesposito0 Going to have our operators try it, though I wonder how well tracking holds up from chest level.

[Vad | Exylos Robotics @vad_shev]

@huggingface Dataset link:
https://t.co/DICQt4wEYJ

[Vad | Exylos Robotics @vad_shev]

A Franka Panda and a WidowX making coffee from the same human demonstration.
New sample on @huggingface : 10 episodes, 14,401 frames, 5 RGB views per robot, 6-DoF object poses.

If you train manipulation policies, tell me what's missing here.

[Vad | Exylos Robotics @vad_shev]

Also relevant for the robotics data as I know right now😄

[Tom Bilyeu @TomBilyeu]

Failure is the most information-rich data stream on planet earth.

[Vad | Exylos Robotics @vad_shev]

@lukas_m_ziegler Went through the repo today and the Quest teleop part isn't in there yet. Waiting for that piece to land to test it properly, mostly curious how much of the desktop engine actually runs natively on the headset.

[Vad | Exylos Robotics @vad_shev]

@rsasaki0109 Useful result for anyone collecting demonstrations through VR teleop. Did the AR cue change how operators actually applied force, or mostly how fast they finished?

[Vad | Exylos Robotics @vad_shev]

Locomotion is solved, they said.
To be fair, everything here is running live in front of a crowd, which is more than most robots are ever asked to do.
Falling over in public is the one failure mode robotics doesn't get to hide.

#WorldHumanoidRobotGames #Robotics

[Vad | Exylos Robotics @vad_shev]

One more sample dataset is on @huggingface

[Exylos @exylos_ai]

Another skill pack is on Hugging Face.
We ran our pipeline on a tougher, multi-step manipulation — making coffee — and it held up cleanly.

Coffee is a real test: handling the cup, working the machine, pouring liquid that behaves like liquid, all in the right order.

We built the dataset to a real customer's spec, so the output lines up with what robotics teams are actually asking for, not our guess at it.

A sample is public on Hugging Face now.
Link in the comments.