Physical AI: End-to-End Learning for Industrial Manipulation
Details
### 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
A short video illustrating the concepts presented in this talk:
https://www.linkedin.com/feed/update/urn:li:activity:7442217801995636737/
***
### References
[1] Zhao et al., Learning fine-grained bimanual manipulation with low-cost hardware, 2023
[2] Chi et al., Diffusion Policy: Visuomotor Policy Learning via Action Diffusion, IJRR, 2025
[3] Yan et al., ManiFlow: A General Robot Manipulation Policy via Consistency Flow Training, 2025
***
### Speaker
Mark N. Bastourous is a Mechatronics and AI Engineer focused on building intelligent systems that interact seamlessly with the physical world. He holds an MSc and a PhD in Computer Science, specializing in AI for robotics.
His experience spans the full lifecycle of electromechanical systems—from CAD design to industrial deployment. He has worked on autopilot systems for marine vessels and autonomous vehicles, as well as precision and uncertainty modeling in industrial robotics.
His research focuses on visual feedback in multi-robot systems to coordinate collective motion across domains, including aerial (UAV), ground (AMR), and marine (USV) platforms.
He is currently an R&D Engineer at IRT Jules Verne.