Raphaël Lafargue (@raflaf_) profile picture

Co-founder @ Asim YC F26 | ex-Sony Researcher | PhD in AI/CV

France
Joined September 2024
Raphaël Lafargue retweeted
Today we're open-sourcing HubEgo-100K: the most diverse egocentric dataset on 🤗 @huggingface. 100,000 hours of skilled tasks from 100,000+ different Hub contributors. Free for research and commercial use 🧵
119
114
42
545
361,262
Raphaël Lafargue retweeted
OpenAI Build Day with the gang 🤝
3
3
1
19
780
Raphaël Lafargue retweeted
we usually put robots to the test. today, it’s our logo’s turn @OpenAIDevs
3
5
358
Raphaël Lafargue retweeted
We’re launching Asim (YC F26) to help robotics teams find failures before deployment. A robot completes a task perfectly. Can it handle an unfamiliar object, reason about how parts fit together, or recover when an interaction goes wrong? We’re building standardized benchmarks for physical intelligence. We test real robots through repeatable scenarios that probe the capabilities behind successful actions: spatial reasoning, understanding physical interactions, dexterous manipulation and adaptation. Teams can identify where performance breaks down, get actionable engineering feedback and measure whether the next software update actually improves it. Building robots? Let’s put them to the test → asim-group.com Asim. Benchmarking physical intelligence.
66
14
7
185
33,482
Your last point, if not addressed properly, could destroy the whole industry! We're working on it now!
What I’m seeing in Physical AI and where I think things are heading. 1. We haven’t had the ChatGPT moment for robotics yet. Models are getting better, but we still don’t have a robot foundation model that can reliably handle a wide range of tasks across different robots. We’re also still figuring out how much data and compute it’ll take to get there. 2. Distribution could become one of the biggest moats in robotics. Companies with robots already working in the real world get access to failures and situations that are hard to recreate in a lab. The faster they learn from those experiences, the bigger their advantage could become. 3. There probably won’t be one winning data recipe. Human demonstrations, egocentric video, teleoperation, robot trajectories, simulation. Different tasks will need different combinations, and figuring out the right mix could matter more than just collecting more data. 4. Deployment infrastructure could become a massive business. Operating thousands of robots brings a completely different set of problems around fleet management, remote interventions, debugging, and maintenance. Existing tools are solving parts of this, but there’s still a lot to build. 5. Simulation is not there yet. Simulation has come a long way, but there’s still a big gap between what works in simulation and what works on real hardware, especially for complex manipulation tasks. A lot more innovation is needed here. 6. Robotics inference will become a much bigger focus. Running powerful models on robots is still expensive and power-hungry. Figuring out what runs on the robot versus on edge or cloud infrastructure could make a huge difference to cost and performance. 7. Specialized robotics companies could scale faster than general-purpose humanoids. Warehouses and factories are much more predictable than homes. I think some of the next big commercial successes will come from companies solving specific industrial problems really well. 8. Hardware and supply chains could decide who scales fastest. Building great models is one thing, but manufacturing robots cheaply and at scale is another. China’s existing hardware ecosystem gives it a huge advantage, and building similar manufacturing capabilities elsewhere won’t be easy. 9. Safety and verifiability could become a huge category. As robots become more autonomous and start operating around humans, we’ll need better ways to verify their decisions, detect failures, and ensure safe behavior. Traditional robotics safety systems alone won’t address every new challenge.
1
3
248
Raphaël Lafargue retweeted
Chasing better physics solvers and new simulation frameworks will get us 90% of the way there, but the only real test of a robot's capabilities, and safety, is in the physical world. ​As this survey by @EliJalbout et al. points out, policies love to exploit simulator corner cases that become dangerous the second they touch real hardware: arxiv.org/abs/2510.20808
1
7
312
Raphaël Lafargue retweeted
See you on Thursday 👀
3
2
12
1,200
Raphaël Lafargue retweeted
ohh, spotted this on my way this morning 👀 Looks interesting!
3
12
337
Raphaël Lafargue retweeted
Thanks for the 1k stars on github!!! My solo project RelateAnything has now crossed 1k stars, only 3 weeks after launch :) RelateAnything is the first Real-Time Open-Vocabulary Relation Prediction model that can run on your laptop. It was trained on more than 19k relation classes and achieve state-of-the-art performance with only 53M parameters! RelateAnything takes any input (boxes, masks etc) and can predict any relation between them (like "person riding horse", "painting hanging from wall" etc) without need for retraining. To celebrate the launch I am releasing 4 new ConvNext weights that push the Pareto frontier even further, as well as a full training recipe code for you to train your own RelateAnything! If you are interested please give it a try: Demo: maelic.github.io/RelateAnyth… Paper: arxiv.org/abs/2609.12552 Code: github.com/Maelic/RelateAnyt… Models and data: huggingface.co/collections/m…
1
3
8
440
So happy to do this!
We’re in YC Fall 2026! 🇺🇸 Two months ago: a hackathon in Paris. Today: co-founders heading to SF We’re building @asim_group to put safety and alignment at the core of robotics. Safe Robots. Fewer people harmed!
3
8
453
🎉 Thrilled to announce our paper "Oops, I Sampled it Again: Reinterpreting Confidence Intervals in Few-Shot Learning" accepted at #TMLR2024 ! 🎉 Big thanks to my amazing coauthors, including @jvlmdr @GriponVincent ! 🙌 🔗 tinyurl.com/yc35wpnt 📽️ tinyurl.com/4k79j49m
1
3
2
14
1,368