Physical AI Research · Roma

TURNING FRONTIER RESEARCH INTO
DEPLOYABLE PHYSICAL AI.

Research-grade solutions to problems off-the-shelf tools and consultants aren't built to solve.

See The Research What We've Built →
Research → Reality
Publications
CVPR · ICLR · ECCV · IROS · AAAI
Collaborations
Berkeley · Stanford · SAPIENZA · Panasonic
Where We Work

Hard problems in perception and physical systems.

Computer Vision

Detection, segmentation and understanding in uncontrolled scenes.

Sensor Fusion

Combining vision, tactile and proprioceptive signals into one state.

Embodied AI

Agents that learn tasks and generalise beyond their training set.

Real-Time Control

Closing the loop on hardware, at the speed the environment demands.

Why Now

The future of AI will not be defined by smarter models, but by systems that work in the real world.

HOW WE GOT HERE READ THE LATEST →
Featured Projects

Research that ships.

ALL PROJECTS →
One-shot imitation learning ItalAI × UC Berkeley

WIZARD

Problem
Industrial robots need re-teaching for every new task and every change in the environment.
Approach
One-shot imitation learning, developed in active collaboration with UC Berkeley: the system watches a single demonstration and adapts to new tasks with zero retraining.
Stage
Active research project, not yet in production.
1
demonstration, then autonomy on unseen tasks — no retraining.
~14×
success rate over the one-shot baseline on our evaluation set — benchmark on the project page.
READ THE PROJECT NOTE
Tactile-reactive manipulation Multi-institutional

T-Rex

Problem
Dexterous manipulation degrades when contact is uncertain and vision alone cannot resolve it.
Approach
Tactile-reactive control fusing touch with vision. A 36-researcher multi-institutional collaboration — ItalAI's contribution is documented in the author list, not claimed as sole authorship.
Stage
Active collaboration, currently in the research and validation stage.
36
researchers across the collaboration — ItalAI among the named affiliations.
2
sensing modalities fused — touch and vision, in one control loop.
READ THE PROJECT NOTE
In action

Research, running in the real world.

Live footage from WIZARD and T-Rex — models operating on real hardware, outside the lab.

WIZARD Wine bottle → cabinet
WIZARD Cup pickup · real world
T-REX Tactile dexterity
BLOG

From research to reality.

READ MORE
Beyond Vanilla SSMs Explainer 02 May 2025

Beyond Vanilla SSMs

Is there an alternative to transformer networks?

ItalAI Research · Sapienza
Transformers
MAMBA · P3
State Space Models, Explained Explainer 02 May 2025

State Space Models, Explained

Mapping the post-transformer landscape.

ItalAI Research · Sapienza
Architectures
MAMBA · P2
Sequence-to-Sequence, Revisited Explainer 02 May 2025

Sequence-to-Sequence, Revisited

Where attention still wins — and where it doesn't.

ItalAI Research · Sapienza
Foundations
MAMBA · P1
What We Build

We build Physical AI systems for hard problems with no off-the-shelf solutions — perception, manipulation and real-time control in environments that break standard tooling.

Every system starts as research: published, benchmarked, and developed alongside the institutions doing the underlying work. That is what it takes to build something that holds up outside the lab.