M4 Control

Efficient and Accurate State Estimation and Feedback Control under Uncertainties

M4 estimates the hidden state of the running process and steers it, even though the process can be observed only partially and with noise.

Illustration of the control task in a forming simulation of a tape laminate. A friction change half-way through forming lets the laminate go slack, and without control the flanges wrinkle. With predictive control, the grippers hold back and keep the tension in band. The plot shows the predicted futures and the one applied.

Work packages

Results of the first funding phase, by work package.

  1. M4.1 Stochastic filtering in large, distributed state spaces

    A Newton-flow particle filter formulates the particle update as an ordinary differential equation derived from a generalised Cramér–von Mises distance. The particles keep equal weights, so the filter avoids degeneracy and global resampling. Deterministic Dirac-mixture samples reduce the number of particles needed.

  2. M4.2 Accurate sensor models for state estimation

    Bayesian neural networks trained with variational inference, Hamiltonian Monte Carlo and Kalman-filter backends are used as uncertainty-aware sensor models. Local calibration tests based on k-d trees, ball trees, Voronoi regions and kernels show where such models are miscalibrated, and new multivariate metrics assess calibration across all outputs.

  3. M4.3 Stochastic feedback control of nonlinear distributed processes

    Sampling-based model predictive control, built on the cross-entropy method and model predictive path integral control, needs no gradients and handles non-differentiable simulators such as Abaqus. Precomputed deterministic samples cover the solution space with fewer samples and give smoother control inputs. The methods were validated on benchmarks and forming simulations.

    Block diagram of sampling-based model predictive control. A sampling-based optimiser draws control input trajectories; each is shot through the system model, and the trajectory costs return to the optimiser. Enlarge figure
    Sampling-based model predictive control optimises control inputs without gradients. Trajectory shooting predicts the behaviour of the system for each sampled control trajectory and can be parallelised efficiently. Figure: KI-FOR 5339.
  4. M4.4 Model-based policy optimisation of nonlinear distributed controllers

    HEPi learns structure-aware distributed controllers in simulation with a heterogeneous graph and SE(3)-equivariant message passing, improving sample efficiency and generalisation across geometries. Learned simulators from M2, such as IGNS and STORM, could serve as fast world models for model predictive control and policy learning.

    HEPi architecture. Equivariant message passing networks process a heterogeneous graph of actuator, object and target nodes, and their outputs are summed into velocity vectors for the actuators. Enlarge figure
    Overview of the Heterogeneous Equivariant Policy (HEPi). Equivariant message passing networks process the observation graph and aggregate the actuator, object and target nodes into the action. Figure: Hoang et al., ICLR 2025.

About the subproject

M4 is designed to build on the learned models of M2 and the settings found by M3. While M3 improves the process from one trial to the next, M4 is meant to keep it on track during a run. Because the forming hardware allows only feedforward actuation, the methods were validated on benchmarks and in simulation.

Principal investigators
Researchers
  • Tai Hoang
  • Xiaozhu Luo
  • Markus Walker
Host institute
Intelligent Sensor-Actuator-Systems Laboratory (ISAS), KIT
People in M4

Publications 13