Seoul National University's RobotValues Exposes 80% Failure Rate in Home Robot AI
RobotValues exposes a critical blind spot: household AI models default to safety and accommodation while systematically ignoring privacy, and can't override those defaults even when explicitly told to
Takeaways
- New benchmark: RobotValues introduces 10,073 image-grounded household scenarios where robots must choose between actions prioritizing different human values.
- Universal bias found: All 10 tested VLMs default to Safety and Accommodation while consistently underselecting Privacy and Security actions across all input modalities.
- 80% failure rate: When told to prioritize a value conflicting with their defaults, models choose the wrong action 83-93% of the time -- accuracy drops 30+ percentage points.
- Not a comprehension problem: Models can identify which value an action prioritizes; they simply fail to override their default preference when selecting between competing actions.
- Fine-tuning fixes it: One epoch of fine-tuning on RobotValues training data improves conflicting-group accuracy by ~50 percentage points, matching larger baseline models.
- Dataset release pending: Data and code will be released upon publication; benchmark is hardware-agnostic and works with any VLM-based robot planner.
Ask a household robot to help an elderly woman who is struggling to walk, and it will almost certainly offer assistance. That sounds fine. But what if she would rather handle it herself? What if calling her husband is the safer option? What if staying nearby but silent is the right call? Today's robot AI systems have no reliable way to reason through these tradeoffs -- and a new benchmark from Seoul National University is the first to measure exactly how badly they fail.
RobotValues is a benchmark of 10,073 image-grounded household scenarios, each presenting a robot with multiple plausible actions that prioritize different human values. The paper's central finding is stark: when you tell a VLM-based robot planner to prioritize a specific value that conflicts with its default preferences, it chooses the wrong action roughly 80% of the time.
The gap nobody was measuring
The robotics field has poured enormous effort into benchmarks for task completion, manipulation precision, and safety compliance. What it has largely ignored is the layer of decisions that happen before task execution -- the moment when a robot must choose between several reasonable actions, each of which trades off a different human value.
Prior benchmarks like BEHAVIOR-1K, RLBench, and ManipBench ask: can the robot do the task? RobotValues asks a different question: when multiple actions are all feasible, which values does the robot actually prioritize? This distinction matters enormously in domestic settings, where robots operate in private spaces and their choices directly affect users' autonomy, dignity, and privacy every day.
The value-conflict problem has been studied in LLMs through text-based moral benchmarks, but it has remained largely unexplored for VLM-based robot planners operating in visual, real-world environments. RobotValues bridges that gap by pairing household images with candidate actions and stakeholder-grounded value annotations.
Building 10K value-conflict scenarios
Constructing the benchmark required solving a hard generation problem: how do you create thousands of realistic household dilemmas where no single action is obviously correct, without the scenarios becoming homogeneous or culturally narrow?

The pipeline runs in five stages, each with automated quality filtering:
- Persona seeding: Each scenario is grounded in a real respondent from the World Values Survey Wave 7 (WVS7), spanning 64 countries. Attributes like age, household size, health, employment, and urban/rural setting are used to generate diverse, realistic household contexts -- preventing the LLM from defaulting to a narrow slice of Western, middle-class scenarios.