Nantes Machine Learning Meetup

### Physical AI for Deformable Object Manipulation

Traditional robotics excels at repetitive, structured tasks but struggles in unstructured environments—such as handling cables, fabrics, or food—where the underlying physics are too complex to model explicitly.
This talk explores the transition toward Physical AI, leveraging end-to-end learning architectures trained from human demonstrations to bridge the gap between simulation and real-world deployment.

***

### Bridging the Gap: ACT, Diffusion, and Flow Matching

We examine three state-of-the-art generative modeling approaches for tackling the high-precision, high-uncertainty challenge of Deformable Linear Object (DLO) manipulation:

• Action Chunking Transformers (ACT) [1]

Learn coherent sequences of bimanual actions.

• Diffusion Policies [2]

Adapt image-generation techniques to iteratively refine robotic trajectories.

• Flow Matching [3]

Provide a fast and stable alternative for continuous motion generation.

***

### From Imitation to Adaptation

By framing industrial manipulation tasks as generative problems, we move beyond rigid programming toward adaptive robotic intelligence.
These models do not simply imitate human behavior—they adapt in real time to slips, tangles, and environmental variations that would typically cause traditional systems to fail.

***

### Preview