robotics
One Transformer Controls Twenty-Eight Plants
A single policy maps short noisy measurement windows to actions without plant matrices or an explicit state estimator.
Summary
A single policy maps short noisy measurement windows to actions without plant matrices or an explicit state estimator.
Classical optimal control normally builds a controller for one plant. Here one transformer imitates linear-quadratic-Gaussian control across linear systems with different dimensions, costs and noise. Standardization, padding, masks and encoded covariance data let the same weights act on every member of the family. Across 28 simulated systems, median cost landed within 0.4 percent of optimum on seen systems; light fine-tuning brought unseen systems to a 1.0 percent median gap. The result is simulation on linear plants, not a guarantee for nonlinear hardware.
Why it matters
A single policy maps short noisy measurement windows to actions without plant matrices or an explicit state estimator.
Limits and context
- The result is simulation on linear plants, not a guarantee for nonlinear hardware.
Key claims
A single policy maps short noisy measurement windows to actions without plant matrices or an explicit state estimator.
Qualification: The result is simulation on linear plants, not a guarantee for nonlinear hardware.
Evidence: source-2026-11-10-009
Sources
- arXiv preprint 2611.11815, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.