Pinned Tweet
🚀 #ICLR2026 Oral 💥
How can we design world models that capture object interactions directly from pixels?
Introducing Latent Particle World Models-the first end-to-end self-supervised, object-centric world model, trained from videos, supporting action/img/lang conditioning.
1/n
How can we design world models that capture object interactions directly from pixels?
Introducing Latent Particle World Models-the first end-to-end self-supervised, object-centric world model, trained from videos, supporting action/img/lang conditioning.
1/n
4
47
282
30,064
Tal Daniel retweeted
🚨Introducing DISCO🪩: Distilled Tree Search over Continuous Actions (Best Paper Finalist🏆 at #CoRL2026)
DISCO combines learned policies with Monte-Carlo Tree Search (MCTS) over continuous actions for long-horizon manipulation tasks.
A thread 🧵👇
DISCO combines learned policies with Monte-Carlo Tree Search (MCTS) over continuous actions for long-horizon manipulation tasks.
A thread 🧵👇
6
30
137
7,831
Tal Daniel retweeted
Robots can’t always see the physics they act under: hidden mass, friction, kinematic constraints. A scalar reward only says "failed." The outcome of an action says much more.
Introducing SCOUT: test-time policy adaptation from action-outcome feedback. Sim-trained, adapts in the es 🧵👇 #CORL2026
Introducing SCOUT: test-time policy adaptation from action-outcome feedback. Sim-trained, adapts in the es 🧵👇 #CORL2026
6
26
126
9,350
Tal Daniel retweeted
What if I told you that Scaled Dot-Product Attention (SDPA) doesn't make a whole lot of sense for ViTs and you can get faster convergence, better performance and near-linear scaling with a few well-steered Gaussians?
Let me introduce you to SSOG (Separable Sum Of Gaussians)!
Let me introduce you to SSOG (Separable Sum Of Gaussians)!
7
60
657
47,671
Tal Daniel retweeted
Kinda relevant to this kind of an object-centric approach: arxiv.org/abs/2603.04553 (ICLR 2026)
1
1
8
378
Tal Daniel retweeted
Boo! 👻 GHOST has been accepted to #RSS2026!
Can we learn manipulation skills from human video (without action retargeting)?
Yes! GHOST learns generalizable manipulation skills by training a hierarchical policy with an embodiment-agnostic sub-goal predictor + troller.
🧵(1/7)
Can we learn manipulation skills from human video (without action retargeting)?
Yes! GHOST learns generalizable manipulation skills by training a hierarchical policy with an embodiment-agnostic sub-goal predictor + troller.
🧵(1/7)
2
14
71
4,628
Self-supervised object-centric learning goes 3D! Get KP, BB and segmentations directly in 3D with 0 annotaions, super useful for robotic manipulation! Amazing work by @EllinaZhang (she will be at #ICML2026 in-person)!
🧩 #ICML2026 💥 How can a model discover the 3D objects in a scene—their shape, color, and position—without any labels? Introducing 3D-DLP, a self-supervised object-centric model that decomposes colored 3D scenes (RGB-D and voxels) into a set of 3D latent particles.
3
7
37
4,019
Tal Daniel retweeted
Excited to announce that our new book Reinforcement Learning Foundations is available for preorder!
W/ @shiemannor and @YishayMansour
30% discount and free shipping till July 21
Discount code 148294
cambridge.org/il/universityp…
W/ @shiemannor and @YishayMansour
30% discount and free shipping till July 21
Discount code 148294
cambridge.org/il/universityp…
3
44
280
15,137
Object-centric representations for tactile data! And for a good cause! Great work!
Excited to present the first-ever 3D, generalizable, handheld tactile imaging system! 👇
Our system learns to visualize the inner structure of soft objects purely through touch -
"More with LESS: Local Scene Representation for Tactile Imaging," has been accepted to #RSS2026 🇦🇺
e the locality of touch for large-scale tactile SSL? Read on! 🧵
0
0
11
584
Tal Daniel retweeted
VLA is 95% certain about current action. Will it 95% succeed in the task?
Obviously, not necessarily. But if you’re clever, you can *calibrate* action prob. to task success.
Our #ICML2026 paper formulates this + SOTA algorithms based on new connection to RL temporal differences
Obviously, not necessarily. But if you’re clever, you can *calibrate* action prob. to task success.
Our #ICML2026 paper formulates this + SOTA algorithms based on new connection to RL temporal differences
Excited to announce our paper is accepted to #ICML2026!
VLA robot policies are capable, but do they know when they're failing? We tackle this with TDQC: a framework for sequential uncertainty calibration, with a surprising link between calibration theory and RL value function.
0
9
47
8,029
I’m attending #ICLR2026 in Rio this week to present LPWM! Friday April 24 Poster Session 3 10:30, Oral Session 4B 3:15p. Happy to chat about self-sup object-centric learning and world models. I’ll be on the job market soon and looking for exciting opportunities!
#ICLR @iclr_conf
#ICLR @iclr_conf
🚀 #ICLR2026 Oral 💥
How can we design world models that capture object interactions directly from pixels?
Introducing Latent Particle World Models-the first end-to-end self-supervised, object-centric world model, trained from videos, supporting action/img/lang conditioning.
1/n
3
8
52
6,375
Tal Daniel retweeted
I will be at #ICLR2026 this week to present our work on Hierarchical Entity-centric Reinforcement Learning!
Come by our poster (Thursday Poster Session 2 P4-#4712) and reach out anytime to talk about #ReinforcementLearning #WorldModels #HierarchicalRL
Come by our poster (Thursday Poster Session 2 P4-#4712) and reach out anytime to talk about #ReinforcementLearning #WorldModels #HierarchicalRL
Learning accurate World Models for long horizon planning is hard.
So what minimal aspect of world dynamics must a model capture to achieve complex goals?
We find a simple and effective solution in our #ICLR2026 paper, which we will present as an Oral at @worldmodel_26.
(1/n)
1
7
51
4,577
Tal Daniel retweeted
🚀 Excited to share ViPRA: Video Prediction for Robot Actions
📍 Accepted to #ICLR2026 @iclr_conf
🏆 Best Paper — #NeurIPS2025 Embodied World Models Workshop
Robot learning today still needs millions of action labeled videos.
Yet videos are abundant — from humans and the web s. Meanwhile, pretrained video models already learn rich dynamics.
ViPRA is a recipe for turning pretrained video models into robot policies while enabling robot learning to scale with actionless videos.
🧵 Thread ↓
📍 Accepted to #ICLR2026 @iclr_conf
🏆 Best Paper — #NeurIPS2025 Embodied World Models Workshop
Robot learning today still needs millions of action labeled videos.
Yet videos are abundant — from humans and the web s. Meanwhile, pretrained video models already learn rich dynamics.
ViPRA is a recipe for turning pretrained video models into robot policies while enabling robot learning to scale with actionless videos.
🧵 Thread ↓
2
41
270
26,239
Join us for live paper (Latent Particle World Models) overview + Q&A today (March 24) at 4pm PT / 7pm ET!
x.com/TalDaniel8/sta…
x.com/TalDaniel8/sta…
🎤 Dr. @TalDaniel8 (@CarnegieMellon ) In a deep dive
w/ @ceciletamura of @ploutosai
How can AI discover objects, model uncertainty, and predict the future from raw video alone?
🔴 [world.ploutos.dev/stream/berserk…](world.ploutos.dev/stream/berserk…)
0
1
7
456
During training, the posterior latent actions condition the dynamics module that predicts the next-frame prior.
A KL regularization term aligns this prediction with the latent policy’s output, forming a VAE-style objective over particle transitions.
7/n
A KL regularization term aligns this prediction with the latent policy’s output, forming a VAE-style objective over particle transitions.
7/n
1
0
8
892
Huge thanks to @DanHrmti, @carl_qi98, Amir Zadeh and Chuan Li (@LambdaAPI), @AvivTamar1, @davheld, and @deepakpathak — incredible collaborators who made LPWM possible.
15/n
15/n
1
0
6
801
🌐 Learn more about Latent Particle World Models (LPWM):
Website: taldatech.github.io/lpwm-web/
Paper: arxiv.org/abs/2603.04553
Code (+ tutorial): github.com/taldatech/lpwm
16/n
Website: taldatech.github.io/lpwm-web/
Paper: arxiv.org/abs/2603.04553
Code (+ tutorial): github.com/taldatech/lpwm
16/n
0
1
10
819