🤖 Understanding what exactly happens during your robot policy deployments is key to improving it. But interpreting raw rollouts is hard because they lack structure.
Introducing Behavior Graphs: turn many rollouts into an interpretable graph of what the policy is doing.
🧵👇
Everyone's been showing Fable and GPT-6 Astra driving robot arms.
We trained an open 27B Qwen model to do it, purely from its own attempts in simulation, with zero demonstrations.
And it can run locally at 220+ tok/s on an RTX 5090.
On a real robot it had never touched, it now succeeds in 18 of 30 tests, up from 0 before training. It also copies a painting and designs its own chemistry experiment.
It generalizes and follows instructions better than π0.5, and on LIBERO it comes close to Astra max.
All without losing its general capabilities, like math and coding.
Built on Qwen3.8-27B. We've open-sourced the model weights, code, and full training infra! Check out our blog at https://t.co/CYHRKigPQP.
Even the simplest weight edit, a deletion, can teach an LLM new skills!
Meet Deletion Tuning (#NeurIPS2026): just zeroing <0.1% of weights matches full fine-tuning's accuracy at scale. It's thousands of times smaller and avoids forgetting, a step toward continual learning. 🧵👇
#CoRL2026 - debate on "The current recipe of behavior cloning pre-training + in-domain fine-tuning will lead to commercially viable generalist robots."
Moderated by Vincent Vanhoucke, who assigned the sides (so not necessarily anyone's personal view). "Generalist" = an end user can re-task it zero-shot, within a well-specified domain. "Commercially viable" = a company builds or runs it profitably.
Pro:
👍 Chelsea Finn: reliability needs lots of in-domain data + optimizing for your objective. like coding models - train on general text and you get a novice coder, train on code and it gets reliable
👍 Pulkit Agrawal: you don't need full generality - get 1% of the pick-and-place market and that's close to a billion dollars in revenue. RL in sim can close the gap on fast, reliable control
👍 Marc Toussaint: the claim is commercial viability, not the best science
Con:
👎 Leslie Kaelbling (of course super energetic lol): "the current recipe" keeps growing - hierarchy, reasoning, world models, coding agents. without a clear problem distribution we can't answer the question, and we probably need a much more heterogeneous architecture. "I think AI people have this kind of disease of romanticism that makes them want to have one true trick." also, language was designed to be learned from very few samples, robot problems weren't
👎 Edward Johns: robot data isn't free - you pay people to collect it, and everyone collects the same data because nobody shares it.
👎 Yonatan Bisk: "I think it's only going to be commercially viable if we get away from having to pay people to collect all of this data." GPT-1 to GPT-3 were weak because they only imitated, and a safety-critical robot has to handle what it's never seen
• another point: covering all the states a robot might reach with demos will never be cheap. and for many tasks the data may cost more than automating the task is worth - cloth folding is one of the few where it pays off
• also, commercially viable means beating competitors, not just breaking even
🗳️ my estimate from the room: a quarter to a third of the room voted pro at the end, the rest con. Vincent: "As far as I can tell, you haven't convinced anyone."
💡 quick thoughts
• it depends on the problem - which tasks, under which assumptions?
• robots don't have internet-scale data, so structure may matter more
I'm giving a talk today at the SPIN workshop #CoRL2026 on Learning by Trying: Physical Reasoning from Puzzles to Contact.
If you are interested in who this person is and in a deep dive on how we unlock creativity and problem solving in robotics, come by!
9:10am in JW Room 208
In robotics, we know all about moving in the world while avoiding contact with it. But what is there is no collision-free path through a messy home or workshop?
MessyNav enables a robot to be deliberate about what things it can push out of the way to clear a path. #Corl2026
@tylerlum23 and I have been busy bringing Play2Perfect to new robots and simulators! Here’s a sneak peek at two robot setups in IsaacLab and MJLab 👇
Come chat with us at our poster today at 5pm at #CoRL2026!
Four posters from our lab at #CoRL2026 Poster Session 1 today, 5–6:30 PM at the JW Marriott 🤖
🟣 #45 MessyMem
🔵 #72 CHORUS
🟢 #120 Mixture of Frames
🟠 #155 MessyNav
Come say hi! Links below 👇
1/8
The real-world is messy. 🤖🧹
If there's no collision-free path, a robot shouldn't just stop, it should move objects out of the way. But which object can it safely move?
Introducing MessyNav: Zero-Shot Navigation via Manipulation✨ Accepted at #CoRL2026 🚀
🧵⬇️
Heading to #CoRL2026 🤠
I'll be presenting MessyMem (https://t.co/iiKgjkCiLM) at the main poster session and as a SISREN workshop spotlight!
📍 Main Poster: Nov 9, 5–6:30 PM | Griffin 2F, #45
🎤 SISREN Workshop Spotlight: Nov 12, 3:05–3:35 PM | JW 306
Hope to see you there! 👋
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?
Doing manipulation with high inertia robots limits fine dexterity.
@younghyo_park and @BronarsToni have built up a solid line of work analyzing the science of robot learning so we can scale more efficiently! Great work 👏
To get an extremely fast system 1 loop, accurately modeling actuation dynamics is critical.
A cool method to calibrate this sim-to-real gap using realworld data!
Robots need memory 🧠: remembering where your favorite cup is or which drawer is locked.
How can a robot remember across large spaces, interactions, & fine-grained detail?
Introducing MessyMem (accepted to #CoRL2026 🚀): Learning-from-Doing Memory for Mobile Manipulation. 🧵
Setting up new tasks with Play2Perfect is super easy!
Here are 2 additional envs + assets that Opus 5 generated from SCRATCH. The same RL training code works without any modifications. Go play with the awesome demo @tylerlum23 created :)
https://t.co/zX81DfApOD
Why do Vision-Language-Action (VLA) models still struggle with contact-rich manipulation? 🤖
Excited to share our new paper: Demystifying When and Why VLAs Fail in Contact-Rich Tasks and How to Fix Them. 🧵 (1/8)
Long horizon bimanual mobile manipulation requires reasoning in many coordinate frames: base, L/R hands, etc. In which frame would policy work best?
It really depends, so don’t pick. Mixture of Frames Policy: denoise in multiple frame in parallel.
🌐https://t.co/GnKSeNqzHA (1/9)