RobotValues: Evaluating Household Robots When Human Values Conflict (arxiv.org)

Researchers have introduced RobotValues, a groundbreaking benchmark designed to evaluate household robots based on their ability to navigate value-conflicting scenarios. Traditional assessments of robots typically focus on task completion, but real-world domestic environments present challenges where robots must prioritize values beyond merely finishing a task, such as upholding human autonomy, efficiency, and social norms. RobotValues consists of 10,000 scenarios, each featuring a realistic household image with various actions a robot could take that reflect different human values, created through a blend of large language model (LLM) assistance and stakeholder input. This benchmark is significant for the AI/ML community as it highlights the critical need for robots to balance competing values in home environments. Initial evaluations using RobotValues revealed that existing vision-language models often default to prioritizing safety and accommodation, while neglecting actions that uphold privacy. Alarmingly, when prompted to make decisions aligned with conflicting values, these models miscalculated their responses 80% of the time. The findings underscore the necessity for a multifaceted evaluation approach that incorporates value-based decision-making alongside traditional performance metrics, thereby guiding future developments in socially-aware robotics.