Generalizable Output Feedback Optimal Control via Transformers Without Explicit State Estimation
Generalizable Output Feedback Optimal Control via Transformers Without Explicit State Estimation
This paper presents a transformer-based policy capable of controlling a family of partially observed, noisy linear time-invariant systems without requiring explicit state estimation or plant matrices. Trained to imitate optimal Linear Quadratic Gaussian (LQG) control across heterogeneous MIMO systems with varying dimensions and noise levels, the approach uses shared representations with standardization and masking. At runtime, the policy generates control actions from a short window of noisy measurements. The study characterizes the relationship between the optimal LQG policy and the static transformer, expressing closed-loop excess cost as the energy of deviation from optimal action. Evaluation on 28 simulated systems shows the trained policy reaches optimal costs within a median gap of 0.4% for seen systems, with fine-tuning reducing the median gap to 1.0% for unseen systems.
The authors demonstrate a transformer-based policy that achieves near-optimal control costs on heterogeneous linear systems without explicit state estimation, validated on simulated benchmarks with quantified performance gaps for seen and unseen system families.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S4
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
d no state estimator. We characterize the relation between the optimal LQG policy (represented as a dynamic state-space model) and a static transformer acting on a finite window of measurements. We then express the closed-loop excess cost of any causal policy as the energy of its deviation from the optimal action. Finally, we evaluate the approach on 28 simulated systems. The trained policy reaches the optimal cost t
Open source S4
o within a median gap of $0.4\%$ for seen systems. For unseen systems, a light fine-tuning step brings the median gap to $1.0\%$.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org