Pinned Tweet
Humans learn by interacting with the world: How can we teach robots to similarly adapt to ambiguous dynamics? Introducing Wiggle and Go! #CoRL2026. We find that a specialized model can learn rope properties from just one interaction, and complete complex dynamic tasks.
The robot gures out how it behaves, then goes, using the estimated dynamics to execute the task. No real-world training data. No retries. #robotics #robotlearning
The robot gures out how it behaves, then goes, using the estimated dynamics to execute the task. No real-world training data. No retries. #robotics #robotlearning
15
31
234
51,707
Really enjoying trying out DeckWerk! I've been surprised how well agents are at helping create slides/animations (without the typical unwanted agent-styling). Adding my own shortcuts to the application where I needed them also felt smooth. Highly recommend people try it out
New release of DeckWerk, the open-source, multi-platform, agent-enabled slide editor! Lots of improvements: Better UI, simplified agent support, better headless mode (=local Google Slides), support for code blocks, more shapes, more animations, better keynote import, ...
Download & install binary here:
deckwerk.org/#install
1
1
3
564
A side note for anyone that does try it out. I found that claude opus was way better than gpt sol at helping edit. Not quite sure why
1
0
1
38
Arthur Jakobsson ✈️ CoRL '26 retweeted
[1/6] What if robots could learn from failures and recover autonomously?
Introducing FAR, a test-time learning framework that enables failure recovery and continual improvement.
No human corrections. No environment resets.
Accepted to CoRL 2026! 🤖
Introducing FAR, a test-time learning framework that enables failure recovery and continual improvement.
No human corrections. No environment resets.
Accepted to CoRL 2026! 🤖
5
9
75
3,985
Awesome work! Can you ask G0.5 "Do you REMMember the 21st night of September?" now?
Introducing REMM: Revisable Embodied Memory for Manipulation.
Robot memory needs to do more than retain past observations. If an intervention undoes an earlier step, the robot also needs to work out which completed conditions no longer hold and what needs to be repaired.
REMM porting visual evidence. When that evidence changes, it revises the affected conclusions and schedules a repair around a frozen G0.5 manipulation policy. Real robot experiments are done on the @GalaxeaDynamics R1Lite.
📄 arXiv: arxiv.org/abs/2610.12710
🌐 Project: revanthsenthil.github.io/remm/
1/8 🧵
#Robotics #RobotLearning
1
0
4
323
Arthur Jakobsson ✈️ CoRL '26 retweeted
Surprisingly a robot can sometimes do highly-dynamic manipulation by just copying the demonstration 🤠. Without any learning, cracking a whip or lassoing a cleat can be done directly from human motion capture. And amazingly this process can be nearly automatically implemented by the full blog post with fun visuals: krishnasuresh.org/blog/2026/robo…
Turn on sound for full effect 🔈
Quick side note: the whips shown in this post are called signal/stock whips. These whips are built to generate the loud cracking noise when the tip breaks the sound barrier and is a popular art form: youtube.com/watch?v=ietBtr….
Turn on sound for full effect 🔈
Quick side note: the whips shown in this post are called signal/stock whips. These whips are built to generate the loud cracking noise when the tip breaks the sound barrier and is a popular art form: youtube.com/watch?v=ietBtr….
33
41
407
113,835
Arthur Jakobsson ✈️ CoRL '26 retweeted
"The robot wiggles a rope once, figures out how it behaves, then goes, using the estimated dynamics to execute the task. No real-world training data. No retries."
Are you getting the idea that a revolution in robotics has started?
Zero shot learning.
Are you getting the idea that a revolution in robotics has started?
Zero shot learning.
Humans learn by interacting with the world: How can we teach robots to similarly adapt to ambiguous dynamics? Introducing Wiggle and Go! #CoRL2026. We find that a specialized model can learn rope properties from just one interaction, and complete complex dynamic tasks.
The robot gures out how it behaves, then goes, using the estimated dynamics to execute the task. No real-world training data. No retries. #robotics #robotlearning
29
27
251
22,760
Last week at ECCV @vincesitzmann said the only way to identify his computer cable's properties for robotics was with a video model. This paper demonstrates that specialized world models still have a place. Though, I think a VWM is an exciting next test for Wiggle-based system identification!
Humans learn by interacting with the world: How can we teach robots to similarly adapt to ambiguous dynamics? Introducing Wiggle and Go! #CoRL2026. We find that a specialized model can learn rope properties from just one interaction, and complete complex dynamic tasks.
The robot gures out how it behaves, then goes, using the estimated dynamics to execute the task. No real-world training data. No retries. #robotics #robotlearning
1
0
16
7,575
Arthur Jakobsson ✈️ CoRL '26 retweeted
There’s something very human about this.
When we pick up an unfamiliar object, we move it around a little before committing to an action.
Here, the robot does essentially that with a rope, turning one interaction into enough information to reason about what comes next.
When we pick up an unfamiliar object, we move it around a little before committing to an action.
Here, the robot does essentially that with a rope, turning one interaction into enough information to reason about what comes next.
Humans learn by interacting with the world: How can we teach robots to similarly adapt to ambiguous dynamics? Introducing Wiggle and Go! #CoRL2026. We find that a specialized model can learn rope properties from just one interaction, and complete complex dynamic tasks.
The robot gures out how it behaves, then goes, using the estimated dynamics to execute the task. No real-world training data. No retries. #robotics #robotlearning
3
2
16
2,153
Arthur Jakobsson ✈️ CoRL '26 retweeted
Introducing PointZero—a 3D world model pre-trained without robots.
Dexterous manipulation requires understanding diverse 3D dynamics, but current approaches rely on expensive robot data. So how can we pre-train a 3D dynamics model without it?
PointZero introduces a simple idea: point tracks yields transferable 3D dynamics!
Our model achieves SOTA results on:
🔥 Zero-shot 3D dynamics
🔥 Action-conditioned 3D dynamics (post-training)
🔥 Imitation learning (post-training)
Details and links 👇 1/6
Dexterous manipulation requires understanding diverse 3D dynamics, but current approaches rely on expensive robot data. So how can we pre-train a 3D dynamics model without it?
PointZero introduces a simple idea: point tracks yields transferable 3D dynamics!
Our model achieves SOTA results on:
🔥 Zero-shot 3D dynamics
🔥 Action-conditioned 3D dynamics (post-training)
🔥 Imitation learning (post-training)
Details and links 👇 1/6
8
46
261
29,080
Arthur Jakobsson ✈️ CoRL '26 retweeted
Had a great time at ECCV last week. Gave my first talk. Met amazing people. Enjoyed good food and wine by the sea. Until next time, Sweden!
5
3
72
2,715
Results in real:
3.55 cm average error on 3D target striking vs 15.29 cm for uninformed baselines
0.95 Pearson correlation between sim and real rope motion on unseen trajectories
3.55 cm average error on 3D target striking vs 15.29 cm for uninformed baselines
0.95 Pearson correlation between sim and real rope motion on unseen trajectories
1
0
6
690
Site + videos: wiggleandgo.github.io
Paper: arxiv.org/abs/2604.22102
With Abhinav, @pkarthik0410 , @_krishnasuresh , @YaoYunchao , @bardienus , @shahram_n95 , and my advisor @jeff_ichnowski. Thank you also to Chris Atkeson for the advice on the paper.
@CMU_Robotics and @SCSatCMU
Paper: arxiv.org/abs/2604.22102
With Abhinav, @pkarthik0410 , @_krishnasuresh , @YaoYunchao , @bardienus , @shahram_n95 , and my advisor @jeff_ichnowski. Thank you also to Chris Atkeson for the advice on the paper.
@CMU_Robotics and @SCSatCMU
0
0
9
682
Arthur Jakobsson ✈️ CoRL '26 retweeted
World-action models typically imagine the future in RGB – but are pixels really the right representation for robotics?
Our bet is no: RGB spends capacity on fine-grained details and variation that are often irrelevant for robot policies.
DINO features, point tracks, and depth tures like semantics, motion, and geometry.
But no single modality captures everything — how can we effectively combine them?
We introduce ✨ModAR✨, which predicts the future one modality at a time, with each prediction informing the next. We train from scratch and find that this formulation performs best.
Our 30.1M scratch-trained model even outperforms a 6B video-model-initialized model finetuned on the same data!
🧵 [1/8]
Our bet is no: RGB spends capacity on fine-grained details and variation that are often irrelevant for robot policies.
DINO features, point tracks, and depth tures like semantics, motion, and geometry.
But no single modality captures everything — how can we effectively combine them?
We introduce ✨ModAR✨, which predicts the future one modality at a time, with each prediction informing the next. We train from scratch and find that this formulation performs best.
Our 30.1M scratch-trained model even outperforms a 6B video-model-initialized model finetuned on the same data!
🧵 [1/8]
12
61
411
38,960
Arthur Jakobsson ✈️ CoRL '26 retweeted
Today we are introducing Atlas! An autoregressive and multimodal DiT for generating image and depth frames. Atlas is a state-of-the-art 3D reconstruction model and outperforms all open-source models. @HaoZhang623 and I had lots of fun pushing the 3D capability of this model; uild with it!
Introducing Atlas:
The world's first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D.
Model the world, move the camera, and simulate space & time.
3
10
112
7,594
Arthur Jakobsson ✈️ CoRL '26 retweeted
How can we get robot hands to “hear” slip and contact through microphones, and react to them?
We’re excited to share VibeAct, an approach that uses piezoelectric microphones embedded in robot fingertips to estimate contact and slip, then learns reactive policies from this s://t.co/7jMLZNpYzp
We’re excited to share VibeAct, an approach that uses piezoelectric microphones embedded in robot fingertips to estimate contact and slip, then learns reactive policies from this s://t.co/7jMLZNpYzp
4
24
144
19,934
Arthur Jakobsson ✈️ CoRL '26 retweeted
Let your robots hear slips with A-SLIP! 🤖🎧
How can a robot detect in-hand slip and estimate its direction and magnitude without cameras or fragile tactile skins?
A-SLIP uses piezoelectric microphones embedded in grippers to hear it.
a-slip.github.io
🧵1/7
How can a robot detect in-hand slip and estimate its direction and magnitude without cameras or fragile tactile skins?
A-SLIP uses piezoelectric microphones embedded in grippers to hear it.
a-slip.github.io
🧵1/7
7
27
190
20,415