Benedek Forrai (@forraibotics) profile picture

co-founder and robotics guy @ mimic | ETH Zürich

Zurich, Switzerland
Joined July 2024
Photo by Benedek Forrai (@forraibotics)Photo by Benedek Forrai (@forraibotics)Photo by Benedek Forrai (@forraibotics)Photo by Benedek Forrai (@forraibotics)Photo by Benedek Forrai (@forraibotics)Photo by Benedek Forrai (@forraibotics)Photo by Benedek Forrai (@forraibotics)Photo by Benedek Forrai (@forraibotics)Photo by Benedek Forrai (@forraibotics)Photo by Benedek Forrai (@forraibotics)
as posted before: every in premier unis, robotics talent is insanely bimodal. you're either goated or you're cooked. tomorrow 14:15 CET we'll be talking at eth's real world robotics course and cover what is in between. public stream link soon
2
2
6
159
*even in lol dont post from bed ppl
1
15
me @ @mimicrobotics with my goulash and tatra tea. not gonna stop. sry all
they really doing everything to twist normal interactions into something sinister
5
260
lmoa good ragebait ultra, walden, etc
If you're not in: -SF -Shenzhen -Bengaluru -Zurich You're NGMI in robotics
3
285
#1 in e2e manipulation - in bech and on the shop floor zero vaporware, zero bs, policies that generalize, hardware that lasts more bangers to come
Today we are previewing FLUX-mimic-x, our new frontier Video-Action Model. We achieve the #1 spot on the RoboLab benchmark among real-time "system 1" models, as well as impressive zero-shot generalization across embodiments including grippers and dexterous hands. FLUX-mimic-x is a 7B-parameter model based on our updated mimic-video recipe and trained on top of FLUX 3 by @bfl_ai. The model is never trained on simulation data or any other data from the RoboLab embodiment beyond that of the original DROID dataset. Our model can be deployed on any robot and any task with minimal adaptation, running on-device and in real-time.
1
16
906
*impossible to teleop on that embodiment w their current stack actually makes sense that they cant do it w a hand: leader follower is super different from how (most) labs do dexterous teleop, and the way they do it makes stuff like this (fast, precise contact) indeed super hard
wtf did he mean by this
101
abundance
Fred again is DJing at the Anthropic roadshow at JPM and you're bearish?
3
183
richard chudton: the blackpill lesson
The bitter lesson says that simple methods that scale well with compute always beat methods based on ad hoc human effort and ingenuity. The blackpill lesson (2026) says that ad hoc human effort and ingenuity is also something that can be produced and scaled with compute.
1
93
another banger from @mimicrobotics . more to come
Today we’re announcing a joint research project with @BMW Group on video-action models for industrial manipulation. We trained FLUX-mimic on automotive tasks using production parts and fixtures: assembly, insertion, kitting and cable routing. On our dexterous robot platform, it achieved success rates of up to 100% and consistently outperformed other models, even with 4x less training data. Collecting task-specific data is the biggest cost of bringing learned policies into factories today. From day one, we’ve believed that working closely with industrial partners on real production data is the only path to deployment.
1
14
504
even discounting human intervention data, deployments build the systems and opex muscle that is needed to dominate the shop floor. if you’re on the frontier but dont deplooooy you’re ngmi
Walden robotics: be full stack, build a good simulation, build a good teleop stack and collect human intervention data at scale.
3
204
KöMaL bros just cant stop winning
Eastern European autism provides full immunity against Chinese honeypot operations.
1
3
281
they solved robotics and realised its joever for corl attendees anyway
This was supposed to be OpenAI robotics booth. Curious what happened?
11
1,619
robotics these days be like
4
19
2
223
8,726
in the next one i need to address the rsi problem
1
3
381
i has already, but to VAMs and not to text interfaces lmoa
btw robotics will fall to the bitter lesson too
4
230
Benedek Forrai retweeted
btw robotics will fall to the bitter lesson too
Replying to @claudeai
Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do. On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.
78
135
24
3,550
261,218
this is so true for any top 1% technical contributor imo not just research never lose that spark brehs
the top 1% PhD: • thinks from first principles • asks questions nobody else thinks to ask • knows how much they don't know • reads papers critically, not religiously • mathematically rigorous • intellectually curious about everything • comfortable being wrong • doesn't confuse citations with impact • knows when to abandon a bad hypothesis • can explain complex ideas to a child • builds things instead of just proposing them • values evidence over authority • learns new fields without asking permission • doesn't chase every research trend • knows which problems are worth solving • collaborates without intellectual arrogance • stays curious after a thousand failed experiments • cares more about discovering truth than publishing papers • remembers why they fell in love with science and somehow still gets excited about a question nobody knows the answer to. anything else?
25
2,279
>larp emergence while overfitting on robots that are not manufactured anymore absolutely based
GPT-6 Astra's robotic capabilities are not emergent. It was trained on robotics specific data. So were GPT-6.1 Sol and Gemini Argon Here's how we know 🧵
4
296
astra fanbois just read this thread b4 we talk thx
"shocking" capability jumps in robotics always happen when someone aligns a robotics skill to another quasi-identical skill that existing frontier models already massively pretrain on for llm-as-robot (astra), it's visual computer use for real time dexterity (VAM/WAM), it's video pretraining
2
351