PhD student in the Robot Learning Lab at @imperialcollege
London
Joined July 2023
- Tweets163
- Following497
- Followers268
- Likes1.1K
Pietro Vitiello retweeted
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.
Pietro Vitiello retweeted
A robot just learned 1,000 distinct tasks in 24 hours, from a single demonstration each. The norm in robotics is the opposite: thousands of demos to master one task. @imperialcollege London researchers inverted the ratio: 31 skills, 42 objects, 1,000 demos in 17 hours. In doing so they challenged the consensus that more data and compute always wins. The trick isn't a bigger model. Here's what they did instead.
Pietro Vitiello retweeted
A robot is supposed to be a tool which one can use to perform a wide variety of tasks. We usually focus on the dream of generalization, but it's interesting instead to focus on that fundamental goal: how can I let someone teach a robot anything I want, quickly?
In "learning a thousand tasks in a day," the authors do just that, using a combination of decomposition and retrieval to get strong generalization with very little data and without training a huge VLA.
Learn more ->
For general-purpose robots to be useful, we must be able to teach them new tasks quickly and repurpose them to new roles. But most techniques for teaching robots — even from human demonstrations — take a great deal of new data. But it turns out that it is possible, in part through decomposing tasks into component subtasks, that it’s possible to learn a robot skills quickly.
@Kamil__Dre @pitvit_ and @Ed__Johns join us to talk about their extensive study on how to teach robots new skills efficiently, while also generalizing to novel object instances, using a combination of task decomposition and retrieval, work that was published in Science Robotics.
Watch episode #94 of RoboPapers now, with @micoolcho and @DJiafei, to learn more!
Thank you for inviting us on the pod @RoboPapers @micoolcho @DJiafei @chris_j_paxton ! I enjoyed the interesting discussions.
If you'd like to know more about what could make your robots more efficient at learning, check this episode out.
Our paper: robot-learning.uk/learning-1…
For general-purpose robots to be useful, we must be able to teach them new tasks quickly and repurpose them to new roles. But most techniques for teaching robots — even from human demonstrations — take a great deal of new data. But it turns out that it is possible, in part through decomposing tasks into component subtasks, that it’s possible to learn a robot skills quickly.
@Kamil__Dre @pitvit_ and @Ed__Johns join us to talk about their extensive study on how to teach robots new skills efficiently, while also generalizing to novel object instances, using a combination of task decomposition and retrieval, work that was published in Science Robotics.
Watch episode #94 of RoboPapers now, with @micoolcho and @DJiafei, to learn more!
Pietro Vitiello retweeted
More folks sld follow the works from @Ed__Johns lab at Imperial. The data efficiency they r going for is incredible.
Their Instant Policy and Learning 1000 Tasks in a Day papers are definitely worth reading.
Photo: taken at Beyond Teleop workshop #icra2026
Pietro Vitiello retweeted
i was visiting a hackathon where 80+ participants were training pi0/0.5, gr00t, smolvla, ACT, DP, etc. on lerobot arms
the best and most sample efficient policies were trained *from scratch*
we still do not have an open source x-embodied GPT-2, but i'm hopeful for this year
Pietro Vitiello retweeted
Here’s the windmill assembly demo we showed at CES 2026 — the one no one saw coming.
North executes a fully autonomous, long-horizon dexterous sequence with sustained hand–eye–tactile coordination and assembly-level precision enabled by tactile feedback.
It’s also robust to disturbance: you can reposition the objects, and North will still identify them and recover the task.
This is powered by CraftNet (VTLA) — using tactile feedback to continuously fine-tune the last-millimeter interaction, enabling reliable execution across 30+ steps.
Read more about CraftNet: sharpa.com/blogs/news/sharpa…
#Sharpa #SharpaWave #SharpaNorth #CraftNet #System0 #CES2026
Pietro Vitiello retweeted
I'm on a singular mission to solve the Physical Turing Test for robotics. It's the next, or perhaps THE last grand challenge of AI. Super-intelligence in text strings will win a Nobel prize before we have chimpanzee-intelligence in agility & dexterity. Moravec's paradox is a curse to be broken, a wall to be torn down. Nothing can stand between humanity and exponential physical productivity on this planet, and perhaps some day on planets beyond.
We started a small lab at NVIDIA and grew to 30 strong very recently. The team punches way above its weight. Our research footprint spans foundation models, world models, embodied reasoning, simulation, whole-body control, and many flavors of RL - basically the full stack of robot learning.
This year, we launched:
- GR00T VLA (vision-language-action) foundation models: open-sourced N1 in Mar, N1.5 in June, and N1.6 this month;
- GR00T Dreams: video world model for scaling synthetic data;
- SONIC: humanoid whole-body control foundation model;
- RL post-training for VLAs and RL recipes for sim2real.
These wouldn't have been possible without the numerous collaborating teams at NVIDIA, strong leadership support, and coauthors from university labs. Thank you all for believing in the mission.
Thread on the gallery of milestones:
Pietro Vitiello retweeted
Robot can now learn 1000 manipulation tasks in a day, and we're still in 2025.
"Researchers at the Robot Learning Lab at Imperial College London recently developed a new imitation learning approach that could allow robots to successfully learn new tasks faster and without requiring substantial training data."
"Using this method, which was introduced in a paper published in Science Robotics, they were able to train a robotic arm to complete 1,000 different tasks in a single day."
Pietro Vitiello retweeted
So excited to finally talk about this work!
Veo is a surprisingly strong world simulator. We fine-tuned Veo on action-conditioned, multi-view robotics data.
Key result: running a policy in the world model is strongly correlated with real-world results.
A few important take-aways:
1) Veo Robotics models real-world physics and robot interactions
2) The base model's world knowledge is retained after fine-tuning and can model OOD scenarios not seen in the robotics data
3) The world model can be used to score task success or failure for a given policy
4) This proves useful for predictive red teaming: simulate dangerous or rare scenarios that would be difficult or irresponsible to execute on the real robot, and judge its performance
I couldn't be more excited about where generalist video models are headed.
Generalist robots need a generalist evaluator. But how do you test safety without breaking things? 💥
🌎 Introducing our new work from @GoogleDeepMind:
Evaluating Gemini Robotics Policies in a Veo World Simulator
veo-robotics.github.io
🧵👇
Pietro Vitiello retweeted
Scalable evaluation is one of the most challenging problems in robotics today. Expensive + slow real-world deployments have been the only gold standard.
But now, generative video models can provide *predictive and scalable* evaluation signal for real-world robots!🌎
Generalist robots need a generalist evaluator. But how do you test safety without breaking things? 💥
🌎 Introducing our new work from @GoogleDeepMind:
Evaluating Gemini Robotics Policies in a Veo World Simulator
veo-robotics.github.io
🧵👇
Pietro Vitiello retweeted
Okay let's talk about this. What happens:
- operator forgets to "clutch out" and tell the robot to stop tracking his hands
- the robot hands go up to its head
- the robot doesnt move its legs to stay balanced - which seems to imply its not doing whole body control like we would expect
- he turns off the robot while its in this unbalanced state, and its arms go to a neutral position close to where the video started-- but because the arms are out of position it makes a huge, fast movement and absolutely smashes that water bottle, sending water spraying everywhere
A lot of people are blaming the operator here but i think forgetting to press a button is, again, pretty human, and you could build a much more robust system here. No reason imo this robot should have fallen.
That’s exactly right, the infrastructure for handling data, training models and setting up robots is immense and the more ambitious the project the worse it gets.
Amazing initiative from @Neuracore_AI and @stepjamUK!
𝗧𝗵𝗶𝘀 𝗶𝘀 𝗲𝘅𝗮𝗰𝘁𝗹𝘆 𝘄𝗵𝘆 𝘄𝗲'𝗿𝗲 𝗴𝗶𝘃𝗶𝗻𝗴 𝗳𝗿𝗲𝗲 𝗮𝗰𝗰𝗲𝘀𝘀 𝘁𝗼 𝗮𝗰𝗮𝗱𝗲𝗺𝗶𝗮.
Imperial’s Robot Learning Lab published some remarkable work last month: teaching a robot 1,000 manipulation tasks in 24 hours from a single demonstration each.
What the paper doesn’t spotlight is the hidden cost behind that result: 5,650+ real-world rollouts and the massive infrastructure required to support them - data collection, sensor sync, format wrangling, storage, and training.
At Neuracore, we see brilliant researchers spending months building pipelines instead of advancing what robots can learn. We see PhD students writing their 15th format converter instead of running their next experiment.
𝗧𝗵𝗮𝘁’𝘀 𝘄𝗵𝘆 𝘄𝗲’𝗿𝗲 𝗼𝗳𝗳𝗲𝗿𝗶𝗻𝗴 𝗳𝗿𝗲𝗲 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺 𝗮𝗰𝗰𝗲𝘀𝘀 𝘁𝗼 𝗮𝗰𝗮𝗱𝗲𝗺𝗶𝗰 𝗿𝗼𝗯𝗼𝘁𝗶𝗰𝘀 𝗹𝗮𝗯𝘀.
Not as a giveaway, but rather 𝗮𝘀 𝗮𝗻 𝗶𝗻𝘃𝗲𝘀𝘁𝗺𝗲𝗻𝘁 𝗶𝗻 𝘁𝗵𝗲 𝗲𝗰𝗼𝘀𝘆𝘀𝘁𝗲𝗺:
• Researchers focus on algorithms, not infrastructure
• Breakthroughs happen faster when data → policy takes days, not months
• Students graduate with battle-tested workflows
• Strong academic validation accelerates industry adoption
Work like this shows the future of robot learning: extreme data efficiency. Our job is to provide the infrastructure that makes discovering the next MT3 frictionless.
If your lab is working on imitation learning, manipulation, or embodied AI, let’s talk.
Credit: @imperialcollege @Kamil__Dre @pitvit_ @vitalisvos19 @Ed__Johns
Paper: robot-learning.uk/learning-1…
Very nice gait. Awesome control @Tesla_Optimus
Pietro Vitiello retweeted
I've been working on deformable object manipulation since my PhD. It was totally a nightmare years ago and my PhD advisor was telling me not to work on it for my own good.
Today, at ByteDance Seed, we are dropping GR-RL, a new VLA+RL system that manages long-horizon precise dexterous manipulation of deformable objects.
This is probably the first real-world RL system to make a robot:
✅ Lace up your shoes end to end
✅ Hit millimeter tolerance repeatedly
✅ Recover from mistakes (See video!)
✅ And complete continuous shoelace threading on a real bimanual platform
📈 Success rate: ↑ from 45.7% → 83.3%
Yes, robots can now actually do this.
Project page: seed.bytedance.com/en/gr_rl
ArXiv: arxiv.org/abs/2512.01801
Pietro Vitiello retweeted
“The thing that happened with AGI and pretraining is that in some sense they overshot the target.
You will realize that a human being is not an AGI. Because a human being lacks a huge amount of knowledge. Instead, we rely on continual learning.
If I produce a super intelligent 15 -year -old, they don't know very much at all. A great student, very eager. [You can say,] ‘You go and be a programmer. You go and be a doctor. Go and learn.’
So you could imagine that the deployment itself will involve some kind of a learning trial and error period. It's a process as opposed to, you drop the finished thing.”
@ilyasut
The @ilyasut episode
0:00:00 – Explaining model jaggedness
0:09:39 - Emotions and value functions
0:18:49 – What are we scaling?
0:25:13 – Why humans generalize better than models
0:35:45 – Straight-shotting superintelligence
0:46:47 – SSI’s model will learn from deployment
0:55:07 – Alignment
1:18:13 – “We are squarely an age of research company”
1:29:23 – Self-play and multi-agent
1:32:42 – Research taste
Look up Dwarkesh Podcast on YouTube, Apple Podcasts, or Spotify. Enjoy!
Pietro Vitiello retweeted
3 years ago we could showcase AI's frontier w. a unicorn drawing. Today we do so w. AI outputs touching the scientific frontier: cdn.openai.com/pdf/4a25f921-…
Use the doc to judge for yourself the status of AI-aided science acceleration, and hopefully be inspired by a couple examples!
Pietro Vitiello retweeted
Learning a thousand tasks in a day
Imagine teaching a robot 1,000 different tricks in a single day—each from just one human demo. That’s the promise behind new work by Kamil Dreczkowski and coauthors: moving robot learning a bit closer to how humans pick up new skills quickly, instead of needing thousands of repetitions per task.
The authors take a very pragmatic route. Instead of one giant end-to-end model, they break manipulation into two simple phases. First, the robot figures out how to line up its gripper with the right object and pose (using language and geometry to find the closest past demo, then classic pose estimation and motion planning).
Second, once it’s in the right place, it replays the fine-grained motion of the original demonstration in the gripper’s own coordinate system. With that recipe, they get a real robot to perform around 1,000 distinct tasks on more than 400 objects, with just one demonstration per task, collected in under a day.
What I like here is the message: you don’t always need a bigger model, you need the right structure. By separating “where to go” from “how to move once you’re there,” MT3 becomes data-efficient, interpretable, and easier to debug—and in the low-data regime, it outperforms standard imitation learning baselines.
For those of us dreaming about robots that set up experiments, handle samples, or reconfigure lab equipment almost on demand, this kind of approach—smart inductive bias plus minimal supervision—looks like a very interesting step forward.
Paper: science.org/doi/10.1126/scir…
