Regret optimal control: EPFL paper offers a metric for measuring robot stupidity

ducha_aiki · x · 2026-09-12

Researcher @duchaaiki explains what "regret" means in control theory via the arXiv paper "Regret Optimal Control for Uncertain Stochastic Systems" (EPFL/ETH): it quantifies loss relative to a clairvoyant optimal policy.

His example: a robot hook moves a 50kg bag of rice — the stupid robot keeps pulling hook and bag, the clever one untangles and proceeds. The paper proposes minimizing regret robustly over randomly sampled system parameters, solvable via semidefinite programming with probabilistic out-of-sample guarantees — effectively a metric for taking "stupidity" into account.

Related event: EPFL research proposes measuring robot 'foolishness' via regret minimization(2 posts)→

Original post →

More from Embodied

Embodied channel →