University of California at Berkeley

Robot Learning, with Inspiration from Child Development

For intelligent robots to become ubiquitous, we need to “solve" locomotion, navigation and manipulation at sufficient reliability in widely varying environments. In locomotion, we now have demonstrations of humanoid walking in a variety of challenging environments.  In navigation, we pursued the task of “Go to Any Thing” – a robot, on entering  a newly rented Airbnb, should be able to find objects such as TV sets or potted plants. The biggest challenges in robotics today lie in manipulation, particularly in dexterous manipulation with multi-fingered hands. Learning approaches have been responsible for recent advances, but they are held up by the lack of “big data” at the scale available in language and vision. I argue that this shortage can be circumvented by taking inspiration from how humans  acquire motor skills in childhood. For dexterous manipulation, multimodal perception is key – vision, touch and proprioception. In my view, visual imitation should be based on 3D/4D reconstruction – then a physics simulator provides a pre-trained world model. The core technology for reconstruction of human bodies, hands, and objects now exists with systems like HMR, HaMeR and SAM 3D.  Visual imitation, while essential, is not sufficient, as policies need to consider contact forces as well. RL in simulation and sim-to-real have been workhorse technologies for us, assisted by a few technical innovations. I will sketch promising directions for future work.

École Polytechnique Fédérale de Lausanne

Performance Boosting for Nonlinear Systems via Stability-Preserving Neural Architectures

Neural network (NN)-based control architectures have driven impressive advances in complex, nonlinear systems, ranging from dexterous robotics to decentralized energy and multi-agent systems. Yet, a core challenge remains: how to fully exploit the expressive power of NN controllers while maintaining rigorous guarantees on closed-loop stability, robustness, and safety.

This talk addresses this challenge by introducing a control framework that enhances performance without compromising stability. We present a complete characterization of stability-preserving policies where the design flexibility is entirely encapsulated within a stable operator. This operator can be represented by broad classes of freely parametrized deep NNs, enabling control design through unconstrained gradient descent or policy distribution sampling.

The proposed framework allows for a rigorous separation of concerns in design: a simple suboptimal regulator ensures stability, while an external optimal loop minimizes a general cost function to optimize performance and safety. This approach naturally extends to a range of settings, including tracking and distributed coordination, while seamlessly incorporating context awareness.

Applications in cooperative robotics and electrical engineering will be discussed to showcase the achievement of complex, high-performance dynamics in realistic scenarios.

University of California at San Diego

Neural Operator Learning for Control

In this talk, we present a novel set of tools and methodologies on physics-informed Neural Operator Control (NOC) for ODE and PDE governed systems. Specifically, we will present NOC for predictor feedback in nonlinear delay systems, NOC for PDE backstepping control, and NOC for optimal motion planning. 

The first part of the talk is about NOC for predictor feedback in nonlinear delay systems. Predictor feedback is effective for delay compensation, yet a critical challenge lies on efficient computation of the predictor operator. We introduce NOC for approximating the nonlinear predictor mapping and prove semiglobal practical stability (dependent on the learning error) of the proposed NOC predictor feedback via back-stepping transformation. 

The second part of the talk is about NOC for PDE control. Model-based methods such as PDE backstepping offer provable guarantees for control but are often computationally prohibitive for real-time implementation. We propose a NOC framework that approximates the mapping from functional coefficients to control gains with desired accuracy. Using neural operator-approximated backstepping gains, we show that our method can accelerate PDE control by up to three orders of magnitude while retaining stability guarantees.

If time permits, we will share our recent advance in designing physics-informed NOC for generalizable motion planning in highly dynamic environments. We propose to encode the obstacle geometries as cost functions and produce fast value function approximations for motion planning, which is defined by the Eikonal PDE. We will conclude the talk by challenges and opportunities in operator learning in control and planning. 

Oxford University

Model-Based and Sample-Driven Synthesis with Logic and Probability 

We are witnessing an inter-disciplinary convergence between scientific areas underpinned by model-based reasoning and by sample-driven learning. In this talk, I will discuss general synthesis problems - of proofs for verification, of controllers/policies for sequential decision making (e.g., in RL), of models and their abstractions - around complex temporal objectives. 

I shall describe how techniques from formal verification (logic, SAT, automata theory) and from learning (neural architectures, sample-driven approaches and probabilistic guarantees) can be together leveraged and integrated to attain both sound and effective synthesis outcomes. 

I will showcase how this general framework is applicable for the formal verification and control of complex dynamical models and of reactive software programs.