Italian Ingenuity in Artificial Intelligence.
Rome
Joined May 2024
- Tweets219
- Following300
- Followers130
- Likes614
🎉 WIZARD is heading to the Compositional Generalization for Real-World Robot Learning workshop at #CoRL2026.
Scale vs. structure: keep the VLA frozen, generate the task-specific adapter around it. LoRA weights from language + one video demo, no test-time optimization.
See you in Austin!
fascetta.github.io/WIZARD/
ItalAI retweeted
We're excited to introduce AW-0: Agentic World Action Models! 🤖 (0/8)
The first World Action Model that plans in code, predicts the future, and controls the robot—all in one generative model.
AW-0 jointly generates:
🧠 Code - explicit skill calls, object arguments, and control flow
🎥 Future video - predictive visual dynamics
🦾 Robot actions - continuous motor control
All three are generated together through joint denoising, letting semantic plans directly shape both predicted futures and robot behavior, with:
🤝 Hybrid execution — generated programs can also run directly through perception and control APIs
⚡ Just 40 ms of additional inference latency over the base WAM on an RTX 5090
📚 Pretrained on ~3,000 hours of robot + human demonstrations
One model for planning, prediction, and execution.
🌐 Project: agentic-wam.github.io/
📄 Paper: arxiv.org/abs/2610.30136v1
💻 Code: github.com/ZhuoyangLiu2005/a…
🤗 Model: huggingface.co/miniFranka/ag…
🧵 How it works—and what we found ↓
ItalAI blog is now live 🚀
First up: WIZARD — predicting LoRA weights for a frozen VLA directly from language + video evidence. Zero-shot execution, no test-time optimization.
The design choices behind the method, it's novelty & real-world applications, as well as the current limits.
🔗 italailabs.com/blog/wizard/
ItalAI retweeted
@_italai has been selected for the Claude Startup Program! 🎉🎉(claude.com/programs/startups)
Thanks @AnthropicAI for enabling more of our AI discoveries.
More to appear here italailabs.com/
From Research to Reality:
Our Co-founder & CTO @GalassoFab10 on what it actually takes to build a successful spin-off 👇
linkedin.com/pulse/from-rese…
Our Co-founder & CTO Fabio Galasso @GalassoFab10 and General Manager Luca Franco @FraLuca17 are at VISMAC today and tomorrow.
If you're attending, feel free to reach out!
#VISMAC26
Meet our new interns 🚀
Stefano is back in the Valley, while Pietro joins for the first time, bringing hands-on robotics experience and a deep interest in LLM interpretability.
As part of our ongoing collaboration with @PanasonicNA R&D labs & @UCBerkeley BAIR, they'll be working on the follow-ups of LEGO and T-Rex.
More to follow!
Great news: T-Rex has been accepted to CoRL 2026 🦖
Most VLA models can't react to touch at the speed contact-rich manipulation actually demands.
T-Rex solves both bottlenecks — data and architecture — with a 100-hour tactile-synchronised dataset and an asynchronous slow-fast architecture that separates visuomotor planning from high-frequency tactile refinement.
Result: 30%+ higher success rate across 12 real-world tasks vs. the strongest baseline.
🧵
The core architectural challenge: vision and touch operate at fundamentally different frequencies. Standard VLA loops are too slow to react to contact as it evolves.
T-Rex separates them — a slow stream for global visuomotor planning, a fast tactile expert running asynchronously, reusing cached visual-language context to respond to force and slip signals in real time without waiting for the full control loop.
Matteo Gioia and Stefano Saravalle worked on T-Rex during their time in the Valley, as part of our ongoing collaboration with @UCBerkeley, @nvidia, @Stanford, and @PanasonicNA.
Everything is open source.
🌐 tactile-rex.github.io
📄 arxiv.org/abs/2606.17055
💻 github.com/ZhuoyangLiu2005/T…
#CoRL2026 #RobotLearning #Robotics #TactileSensing #EmbodiedAI
Back from the Valley:
Matteo and Stefano on T-Rex /🔗tactile-reactive-dexterous.g…, tactile perception, and what it takes to work at the frontier of physical AI.
Read the full Q&A 👇
linkedin.com/pulse/frontier-…
Meet WIZARD 🤖
Given a language prompt and a single unlabeled demo video, it generates the LoRA adapter weights a frozen VLA needs to perform a new task — in one forward pass.
No action labels. No finetuning. No gradient steps at deployment.
The WIZARD paradigm: video in, policy parameters out.
Paper & demos: fascetta.github.io/WIZARD/
T-Rex is finally out in the open 🦖
This is the project Matteo Gioia and Stefano Saravalle have been working on during their time in the Valley as part of our ongoing research collaboration with @UCBerkeley.
T-Rex addresses both the data and model bottlenecks of tactile-reactive dexterous manipulation through tactile-grounded mid-training and an asynchronous slow-fast control architecture for high-frequency tactile refinement.
The result: over 30% higher average success rates across 12 real-world contact-rich manipulation tasks.
Tactile feedback is becoming a key capability for next-generation robotic foundation models.
The thread is well worth a read ↓
Excited to share T-Rex: Tactile-Reactive Dexterous Manipulation 🦖🤖
Touch is fundamental to human dexterity, yet most Vision-Language-Action (VLA) models either ignore tactile feedback or lack the ability to react to high-frequency contact signals.
In this work, we tackle both the data and architectural challenges of tactile-reactive dexterous manipulation.
🦖 A 100-hour tactile-synchronized dexterous manipulation dataset with 7,700+ trajectories, 22 motor primitives, and 200+ everyday objects.
🦖 A tactile-reactive MoT architecture with spatial-temporal tactile encoding and asynchronous high-frequency tactile refinement.
🦖 A scalable training recipe combining 22,889 hours of human egocentric pretraining with tactile-grounded robot mid-training.
Across 12 real-world contact-rich manipulation tasks, T-Rex achieves over 30% higher average success rate than the strongest baseline.
We are fully open-sourcing the dataset, models, teleoperation stack, training code, and inference pipeline.
🌐 Project: tactile-rex.github.io/
📄 Paper: arxiv.org/abs/2606.17055
💻 Code: github.com/ZhuoyangLiu2005/T…
🤗 Dataset: huggingface.co/datasets/zeka…
🧵 Thread ↓
Thanks to everyone who joined us at the MUV Workshop at @CVPR !
@muv_workshop focused on selective data removal,
concept erasure, debiasing & model editing, as well as safe image/video generation & post-training control.
Great to see the community come together to advance the future of safe, privacy-aware, and trustworthy AI.
#CVPR2026
ItalAI retweeted
Everything is ready for MUV @CVPR 2026! 🧠❌
Join us in Denver on June 3 for the Machine Unlearning for Vision Workshop.
📍 Room 1AB
🕒 Afternoon session
Program, speakers, and details:
machine-unlearning-for-visio…
#CVPR2026 #MachineUnlearning #ComputerVision #AI
We’re still #hiring an Applied AI Engineer to join our team 🔊
You’ll work on AI systems for code understanding, automation, and developer workflows — building and evaluating LLM/agent-based pipelines that operate reliably in real-world production environments.
MSc required, PhD preferred.
📩 Send your CV + GitHub / projects → [email protected]
Yesterday our team joined #RaceForTheCure by @komenitalia alongside 200,000+ people supporting the largest worldwide initiative against breast cancer.
Proud to contribute in our own small way and share the experience together with family and friends 🏃🏻♂️💗
