
Dominique Paul | 📍 CoRL
@DominiqueCAPaul • 9,435 subscribers
Founder of Dream Machines @dm_robots, building ML models for robotic arms. Previous: Stats at @ETH & @NeurIPSconf. I post about RL and VLA post-training.
Shorts
Videos

10 PM in Huaqiangbei, SZ. We’re on our way getting dinner with Fateen Anam Rafid and Aurel wonders if anyone could swap the battery in his old iPhone. First stall we ask does it on the spot in 10 minutes. 20€ for a third-party battery, 40€ for an original Apple one.
Dominique Paul | 📍 CoRL85,396 Aufrufe • vor 26 Tagen

Europe just overtook the US in the number of git pushes.
Dominique Paul687,914 Aufrufe • vor 9 Monaten

Two months of ablating π0.5 finetunes on a real manufacturing task. The policy now succeeds 98% of the time. We're publishing all results, all the data, and every run, including the ones that went nowhere. What we learned: 1/ 5x more data took us from 63% to 76%. Pure scaling was the weakest lever we found. 2/ Diversity got more out of the same hours. 4h spread across five scenes beat 4h in the eval scene by 30pp. 3/ Quality did even better. One clean, slower hour on top of the 21h: 76% → 90%. That single hour added more than the 17 hours before it. 4/ So we tried quality on its own: 1h of clean data plus 240 rollouts with human interventions. 1.7h in total, 28% → 88%. That's 12 points above the 21h model with a twelfth of the data. 5/ Inference settings add more on top, with no retraining. They took the 21h model from 76% to 93%, and the 21h+1h model from 90% to 98%. Just as useful: what didn't matter. Learning rate, batch size, relative vs. absolute joints, image augmentation. All less important than we expected. Full write-up linked below.
Dominique Paul82,274 Aufrufe • vor 1 Monat

Around 3 million Chinese private companies face a generational handover in the next 10 years, and it‘s the founders' kids who grew up in the digital age we should probably be watching. Today, we visited a mattress factory in Foshan where the son is taking over, and he's gone all in on the brand: > built an entire museum of sleep explaining the production process > added AI point-cloud scans to match mattresses to body types > or my favourite: wrote a song about the company's history and sang it himself (yes, this was pre-AI slop; last video). Reminds me of many other new owners on TikTok rethinking online marketing. Next to highly automated consumer lines, they sell handmade mattresses filled with white horsetail hair that go for up to ¥600,000 (~€80k, second pic). Horsehair in mattresses is an old European high-end tradition. Tail hair is longer and thicker than mane, so it springs back better and lasts for decades. White hair is finer and rarer and so in part it’s about selling the idea of luxury that some people care about.
Dominique Paul38,672 Aufrufe • vor 28 Tagen

Lighting differences can make a huge difference in robotics. Today, I found a quirk in my model exemplifying this. > I collected 10h of training data. > 3h in, I notice that the left arm following the right arm for the final movement could be good for the final insertion subtask. > I change behaviour. > But the first 3h didn't include much nighttime data (=dark room). 🎯Result The same policy reliably exhibits different behaviour on exactly this subtask depending on the lighting. So many microlearnings to be made in robotics.
Dominique Paul97,968 Aufrufe • vor 5 Monaten

China, day 7 Highlight today was the LED manufacturer and walking their production floor. Comments under my other post today explained that wages in places like these are well above minimum wage, more like ¥6-8k a month. That puts cost per worker close to the factories I visited in Poland … but with roughly 20 more hours of work per worker per week. The longer I'm here, the more China reminds me of the US in one way: the media covers the extremes, and in everyday life you feel very little of them. Key difference: US has more influence over global media while reporting on china abroad is one-sided with few people setting foot here. Lunch was Cantonese and Beijing food, including Peking duck. Very fatty skin, very good. Like with other dishes here, part of the point is showing how fresh it is, so the duck gets carved in front of you. Three tables = three cooks carving. Then the Foshan hydrogen lab: 140 researchers on fuel cells and ammonia-hydrogen. Nanhai, a district of Foshan, is a hydrogen cluster with 160+ companies. One of their projects is powering robots with hydrogen. They also spin off ~10 companies a year. Interesting learning: Research institutes like this one align their planning with China's 5-year plan. Obvious in hindsight. "5-year plan" sounds very communist, but the more I thought about it I thought that an election cycle isn’t that different, except that the guys doing the 5-year plan think more long term. Evening in Foshan's city centre. A temple, a local hotpot cooked in rice water instead of oil or broth. Then walked around the city centre. Fewer towers than Guangzhou, but a surprising number of high-end stores. Could have been Japan. Prices are the same as in Europe. Contrasted with this morning’s factory wages, I am wondering what the average customer in these stores does for a living.
Dominique Paul18,373 Aufrufe • vor 29 Tagen

A few months ago I realised how when you’re building alone you miss out on the random photos other people take of you. Like the ones that pop on your phone or somewhere else later a remind you of a specific phase of your life. So I decided to start taking a few myself, even during the bland stretches. Today, while reading the RL book, I thought of that and recorded a tiny time-lapse. I then later noticed I should have done so during an easier section of the book because I apparently spent 11 minutes on two pages and don’t flip a page until the end 😂
Dominique Paul90,241 Aufrufe • vor 11 Monaten

Our policy is getting pretty good. 88% success rate on actuator unboxing.
Dominique Paul35,667 Aufrufe • vor 4 Monaten

We built a golden robot retriever! > SO-100 attached to a Unitree Go2 > Trained an ACT policy to grab objects that (mostly) also works for OOD objects and backgrounds > System is controlled by an agentic loop (OpenAI) evaluating surroundings & triggering retrieval with the arm > The quadruped is teleoperated Thank you, Valentin Hartmann, Wen, and Brian, for the amazing time together! Thank you Arnie Ramesh @rold_si Pascal Bertrand Lucas Beyer (bl16) OpenAI mimic Elvis Nava Loki Robotics Antonio Arbues for an amazing hackathon!
Dominique Paul118,211 Aufrufe • vor 1 Jahr

Thoughts from a teleop session today: 1/ Teleop is painful. UMI-style grippers are making more and more sense to me: shorter per-episode execution, more data, and more intuitive for factory workers who will be the ones with my product eventually. Wondering if PI resists this because researchers aren't the ones collecting the data? 2/ The take-away of the HF shirt folding post stuck with me: data quality matters most of all. I'm 30 minutes into this task and still making mistakes with teleop. What’s the perspective for non-roboticists? Maybe VR headset is better. Want to try that next. 3/ I’m noticeable better at teleop even when I’m just 15cm closer. 4/ Double-close gestures for re-record (left) and early episode end (right) are a game changer. Credit Andreas Köpf. 5/ Want to gamify my own collection more: thinking of a daily target dashboard. 6/ I’d like to rate each episode with a 1-5 data quality score. Don't wanna throw away bad data away, but still be able to filter top-quality. Maybe possible with foot pedals?
Dominique Paul38,065 Aufrufe • vor 7 Monaten

Today and tomorrow are data collection power days: 465 today with a goal of 1100 new samples total. The Hugging Face recording script previously took 30-40s to process each 15s recording, but I changed it to batch process and now collection is way faster (and more fun)!
Dominique Paul39,273 Aufrufe • vor 1 Jahr

Another reason I’m interested in UMI style grippers is that tasks that‘d take 10s for a human take 5x as long with teleoperation. The lack of haptic and force feedback makes it hard to do movements relying on them like here, passing an object from one hand to another. Adding to that: teleoperation means you’re standing a meter away from the objects you’re operating on, and here for example, the box is obstructing the view of the PCBs. Thank you for great teleoperation Felix Neumann 🤝
Dominique Paul15,434 Aufrufe • vor 10 Monaten

Today I fixed the follower arms to a profile bar and started experimenting with new leader–follower setups. This speaker desk I found on Amazon is the first of three. I like the laptop‘s position in the bottom tray: Doesn’t block view and still accessible for quick code changes. What I still haven’t figured out is the ideal horizontal distance between both the leader and follower arms to make teleop feel more natural.
Dominique Paul10,612 Aufrufe • vor 11 Monaten

First successes in training the robot policy for chess movements! 🧵
Dominique Paul13,474 Aufrufe • vor 1 Jahr
Keine weiteren Inhalte verfügbar