Tensor Networks for Robot Learning

During my PhD at EPFL (2019–2024), supervised by Dr. Sylvain Calinon at the Idiap Research Institute, I pioneered Tensor Networks (aka low-rank tensor approximation techniques) as a powerful computational paradigm for robot learning. These techniques, heavily used in physics and quantum computing, offer interpretability, algebraic manipulation, and can model complex functions with high accuracy – making them particularly suited for robotics.

I developed fast and memory-efficient algorithms for robot exploration, learning from demonstration (LfD), reinforcement learning, and motion planning, with a particular focus on robotic manipulation tasks.

What are Tensor Networks?

Tensor Networks are powerful function approximation techniques that decompose high-dimensional tensors into networks of smaller, lower-rank tensors. They are particularly interesting for robotics as they can break the curse of dimensionality while maintaining high accuracy.

What can you do with Tensor Networks?

Motion Planning: Global Optimization + Multiple Solutions

Control of Under-actuated Systems

First GIF
Pendulum Swing-up

Second GIF
Cart-Pole Swing-up

Control of Systems with Hybrid Contact Dynamics

Control of Systems with Hybrid State and Action Space

First GIF
Object Pushing Task

Second GIF
Real-world

Selected Publications

S Shetty, T Xue, and S Calinon, “Generalized Policy Iteration using Tensor Approximation for Hybrid Control”, ICLR 2024. [Spotlight Paper, 5% acceptance rate] – Website
S Shetty, T Lemobono, T Loew, and S Calinon, “Tensor Train for Global Optimization Problems in Robotics”, IJRR 2023. [Best Paper of the Year Award from IEEE RAS TC on Model-based Optimization for Robotics] – Website
S Shetty, J Silverio, and S Calinon, “Ergodic Exploration Using Tensor Train: Applications in Insertion Tasks”, IEEE T-RO 2021. [Idiap Paper of the Year 2021] – Website
T Xue, A Razmjoo, S Shetty, and S Calinon, “Robust Manipulation Primitive Learning via Domain Contraction”, CoRL 2024.
T Xue, A Razmjoo, S Shetty, and S Calinon, “Logic-Skill Programming: An Optimization-based Approach to Sequential Skill Planning”, RSS 2024.
A Razmjoo, T Xue, S Shetty, and S Calinon, “Sampling-Based Constrained Motion Planning with Products of Experts”, IJRR 2025.
T Xue, A Razmjoo, S Shetty, and S Calinon, “Robust Contact-rich Manipulation through Implicit Motor Adaptation”, IJRR 2025.

Suhan Shetty