arXiv cs.RO·· 2 days agoAI score67
Paper finds sampling-based planning exploits human-preference rewards, proposes χ²-regularized MPPI correction
Summary
The arXiv paper "Goodhart in the Planner" measured how sampling-based control with rewards learned from human preferences performs on the true objective: on HalfCheetah, Hopper, and Meta-World's drawer-open.
This page provides a summary and source link only. Read the full article at the source.
Source: arXiv cs.RO · arxiv.org
Robot models and learningSafety and reliabilityResearch papers#Research#Robot models and learning#Safety and reliability
Part of the event: Study: Preference Reward Planners Harm True Objectives