RobotValues: Value Conflicts in Home Robotics

• RobotValues is a benchmark that evaluates household robot planners by presenting realistic domestic decision points where actions trade off distinct human values.
• The benchmark uses 10,073 image-grounded scenarios and dual task settings (default and value-conditioned) to assess both spontaneous and steerable model behaviors.
• Empirical results show models default to safety and accommodation while struggling to prioritize values like privacy and security when conflicts arise.

Searching arXiv for the RobotValues paper and closely related recent robotics-value-evaluation work. tool call: arxiv_search(query="RobotValues Evaluating Household Robots When Human Values Conflict", max_results=5) RobotValues is a benchmark for evaluating household robot planners in situations where plausible actions trade off distinct human values rather than merely maximizing task completion. It is introduced as a large-scale, multimodal testbed of 10,073 image-grounded decision points in domestic settings, where each instance couples a realistic household image and compact text context with multiple candidate robot actions that prioritize different values. The benchmark is designed to expose a gap in prevailing evaluation practice: a robot may complete a task or satisfy a generic safety criterion while still choosing the wrong action when autonomy, privacy, efficiency, accommodation, or related norms conflict. In the reported evaluations, vision-LLMs used in robotics exhibit consistent default preferences and substantial difficulty obeying explicit value-prioritization instructions when those instructions conflict with their defaults (Han et al., 2 Jun 2026).

1. Benchmark objective and problem setting

RobotValues targets a class of domestic decision problems in which a robot reaches a single intervention moment and must choose among several feasible next actions, each of which serves one human value while compromising another. The benchmark therefore shifts the unit of evaluation from success-only assessment to value-sensitive action selection. The central question is not simply whether a robot can act effectively, but which value it serves when several values are simultaneously at stake (Han et al., 2 Jun 2026).

The benchmark defines two task settings. In the default choice setting, a model receives an image, compact textual context, and candidate actions, and is asked to select the “most appropriate” action. In the value-conditioned choice setting, the same inputs are augmented with a target norm such as Privacy, and the model must choose the action whose annotated value matches that target. This separation distinguishes spontaneous model preferences from steerable value-conditioned behavior.

A recurrent source of confusion in domestic robotics is the assumption that correct task execution or generic helpfulness is an adequate proxy for value alignment. RobotValues is constructed specifically to test cases where this assumption fails. A privacy-protecting action may delay intervention; an accommodating action may reduce efficiency; a safety-oriented action may override autonomy. The benchmark operationalizes these conflicts as explicit choice problems rather than as failures of task specification.

2. Value taxonomy and conflict structure

RobotValues is grounded in a robot-norm taxonomy attributed to Li et al. (2019), comprising 10 categories:

Category	Definition
Safety	“Protect humans from danger.”
Consideration	“Consider human feelings.”
Privacy	“Protect human privacy.”
Security	“Safeguard sensitive information.”
Efficiency	“Complete the given task efficiently.”
Compliance	“Obey social rules.”
Command	“Follow the owner’s commands.”
Accommodation	“Accommodate human behavior.”
Honesty	“Tell the truth.”
Loyalty	“Maximize the owner’s interests.”

Conflicts arise because multiple candidate actions can be simultaneously feasible at a single intervention moment, yet each action advances one norm while stepping on another. The benchmark also augments these coarse labels with fine-grained, stakeholder-grounded values extracted from first-person stakeholder reactions, such as “preserving P1’s personal autonomy by waiting for permission” (Han et al., 2 Jun 2026).

The reported scale and annotation breakdown are central to the benchmark’s characterization. RobotValues contains 10,073 image-grounded decision points, approximately 10K, and 69,134 candidate actions, approximately 17 per scenario before de-duplication. Over all actions, the distribution of robot-norm annotations is: Safety 27.1%, Accommodation 22.1%, Consideration 14.8%, Privacy 8.1%, Efficiency 7.8%, Honesty 7.0%, Compliance 5.4%, Command 3.6%, Loyalty 2.7%, and Security 1.4%. A post-hoc mapping to Schwartz’s basic values yields Security 37.3%, Benevolence 28.0%, Conformity 17.1%, Self-Direction 10.8%, and other categories. This distribution matters because later model behavior is compared against dataset frequency rather than against an assumed uniform prior.

A plausible implication is that RobotValues tests both coarse normative preference and more contextualized stakeholder-sensitive judgment. The coexistence of fixed norm labels and extracted fine-grained values allows the benchmark to separate categorical preference from scenario-specific justification.

3. Data construction pipeline

RobotValues is produced through a multistage, LLM-assisted generation and filtering pipeline intended to preserve household realism while maintaining explicit value trade-offs (Han et al., 2 Jun 2026).

The pipeline begins with persona and context seeds. Persona seeds are drawn from World Values Survey Wave 7 (WVS7), sampling 90,313 real respondents via weighted priority sampling. Each seed supplies attributes including country, household size, age, sex, health, and employment. Scene diversity is enforced through round-robin room type across 10 categories and time of day across 5 slots.

In scenario generation, an LLM such as DeepSeek-v4, GPT-5-mini, Qwen3-235B, or GPT-OSS is prompted to write a realistic household scenario containing a robot task, a description of the intervention moment, and a list of stakeholders denoted P1, P2, and so forth, plus R1, with roles and presence flags. A binary quality-control judge then checks persona fidelity, realism, coherence, and stakeholder materiality.

In candidate-action generation, the system defines 17 “value seeds” by merging eight HRI value-compass categories, attributed to Abbo et al. (2026), with the ten household norms. The LLM generates one action per seed, each with a concise action description and a short justification naming the concrete benefit and trade-off. A judge filters actions for feasibility, plausibility, and safety, and merges near-duplicates.

In stakeholder reaction and value extraction, the pipeline generates, for every action and every stakeholder, a first-person reaction with label in $$\{\text{support}, \text{oppose}, \text{mixed}, \text{neutral}\}$$. Another prompt extracts a single fine-grained prioritized value from those reactions, grounding it in concrete stakeholder concerns. A judge checks that the action genuinely prioritizes the extracted value and that reactions support that interpretation.

In image generation, the system produces a “snapshot” description with fields viewpoint_type, viewpoint, visible_scene, and decision_evidence, and then uses GPT Image 2 at 1280×720 to synthesize a photorealistic egocentric household image while hiding any visible robot embodiment. A judge verifies scenario grounding, physical realism, artifact-free humans, realistic viewpoint, and absence of robot hardware.

At the corpus level, the pipeline starts from 16,000 seeds and, after five stages of filtering, retains 10,073 scenarios, corresponding to 63.0% overall acceptance. This staged retention is important because the benchmark is not merely synthetic text generation; it is synthetic scenario construction with repeated automated adjudication.

4. Formal evaluation methodology

RobotValues evaluates both default value preference and instruction-following under value conflict. These are formalized through Bradley–Terry preference modeling and value-conditioned accuracy (Han et al., 2 Jun 2026).

For default preferences, each selected action is converted into pairwise wins. If an action $a$ with value $i$ is chosen over actions $b$ and $c$ with values $j$ and $k$, then the pairwise counts are incremented as $c_{ij} \mathrel{+}= 1$ and $c_{ik} \mathrel{+}= 1$. A Bradley–Terry model is then fitted with preference probability

$P(i \succ j) = \frac{w_i}{w_i + w_j},$

where the worth parameters $a$0 are estimated by the minorization–maximization algorithm with a weak symmetric pseudocount of 0.5 to avoid zeros. Preference is summarized by the centered log-worth score

$a$1

A positive $a$2 indicates that value $a$3 is preferred above the average.

For value-conditioned evaluation, each scenario is queried once per target norm, and results are aggregated over five random restarts by majority vote. Accuracy is defined as

$a$4

Results are stratified by whether the target norm matches the model’s default preference, conflicts with it, or occurs in a tie. This matched/conflicting/tie decomposition is methodologically important because raw accuracy alone would obscure whether errors are caused by arbitrary confusion or by systematic default-value bias.

The benchmark evaluates the following vision-LLMs: Qwen3-VL-2B, Cosmos-Reason2-2B, Cosmos-Reason2-8B, Molmo2-8B, Molmo2-ER, RoboBrain2.0-7B, InternVL3-2B, InternVL3-8B, InternVL3-3.5-8B, and RLDX-1-VLM. No closed-source “generalist” VLMs such as ChatGPT or Gemini are included. All 10,073 instances are used for evaluation. For a value-conditioned adaptation pilot, the paper reports 11,942 training examples, one epoch of fine-tuning, and a held-out split of approximately 1,131 instances. A real-camera pilot uses two images from an SO-101 wrist camera with Claude Opus 4.7 for scenario synthesis and a fine-tuned Qwen3-VL model for inference.

5. Empirical results: default preferences and failures of override

The central empirical finding is that the evaluated models exhibit stable default value preferences and struggle to override those defaults when explicitly instructed to prioritize conflicting values (Han et al., 2 Jun 2026).

Under the household-norm taxonomy, the default-choice Bradley–Terry extremes place Safety at the top, with centered log-worth scores from +0.38 to +0.70, and Accommodation or Consideration also high, with Accommodation in the range +0.33 to +0.43. At the bottom are Privacy, from −0.45 to −0.94, and Security, from −0.64 to −0.95. These rankings persist under modality ablations using text-only, image-only, or actions-only inputs, indicating that the broad preference ordering is not an artifact of a single modality.

In the value-conditioned setting, accuracy is much higher when the requested norm aligns with the model’s default than when it conflicts. The reported ranges are: Matched 40.2%–51.3%, Tie 12.8%–22.5%, and Conflicting 6.9%–16.8%, with a Matched-minus-Conflicting drop of approximately 30–41 points. The paper summarizes this as models choosing incorrect actions for 80% of the time when instructed to prioritize specific values that conflict with their own preferences. In the expanded findings, the same point is expressed as success only 7%–17% of the time, corresponding to failure approximately 80%–93% of the time.

The benchmark includes an action–value matching ablation that isolates whether models can recognize which value an action promotes. When asked only “Which norm does this action promote?”, Conflicting accuracy rises to 35%–49%, and the drop narrows to approximately 8%–17%. This is a crucial diagnostic result: the models often know which action corresponds to which value, yet still fail to choose it when the requested value conflicts with their default preference. The benchmark therefore attributes the main failure not solely to semantic misunderstanding, but to an inability to act on explicit value-conditioning.

In the fine-grained values setting, Conflicting accuracy is somewhat higher, approximately 21%–30%, but the drop remains large at approximately 23%–33%. This suggests that more contextualized stakeholder-grounded values modestly help, yet do not eliminate default-bias effects.

6. Interpretation, design implications, and scope of extension

RobotValues frames these findings as evidence that household robot evaluation should measure not only task completion or safety compliance, but also whether robots can choose among plausible actions when human values conflict (Han et al., 2 Jun 2026). The benchmark’s interpretation is that current vision-language planners tend to default toward Safety and Accommodation, or sometimes Consideration, while underselecting Privacy- and Security-promoting actions relative to their dataset frequency. The reported underselection is especially notable because Privacy accounts for 8.1% of annotated actions and Security for 1.4%, yet both appear at the bottom of default preference rankings.

The paper offers several explanations for why Privacy is underselected. Vision-LLMs appear to internalize physical safety and task efficiency as more urgent and more straightforward than privacy concerns, which are described as more abstract and context-dependent. Household scenes in pretraining are said to illustrate “helpful” interventions more often than “hands-off” policies, biasing models toward intervention. Privacy also often requires inaction or careful waiting, which such models are less likely to choose under default helpfulness heuristics. These are presented as interpretive explanations rather than as directly measured causal mechanisms.

For household robot design, the benchmark recommends that evaluation suites include value-conflict benchmarks, that practitioners use value-conditioned prompting but validate with held-out Conflicting tests, and that supervised fine-tuning on synthetic value-conflict data be considered while monitoring trade-offs between Matched and Conflicting accuracy. It also recommends stakeholder-in-the-loop feedback. The paper further states that hand-coded rule systems or simple prompt injections are insufficient, and that learning or planning architectures should incorporate explicit value representations and refinement steps such as inner monologue or reflective planning.

The benchmark also reports that fine-tuning on RobotValues can improve value-conditioned accuracy, with an example gain of +50 percentage points in the Conflicting setting, though with a slight reduction in Matched accuracy. The real-camera pilot suggests some transfer to physical setups, but the evidence reported is intentionally limited: it involves two images from an SO-101 wrist camera rather than a broad real-world deployment study. This suggests cautious interpretation of transfer claims.

RobotValues is described as the first large-scale, multimodal benchmark explicitly testing household robots in realistic value-conflict scenarios. Possible extensions named in the paper include broader value taxonomies such as Schwartz’s and cross-cultural surveys, dynamic multi-turn dialogs for clarifying uncertain value priorities, additional modalities such as audio cues and haptic signals, real-robot trials with on-device inference and online updating, multi-agent scenarios with conflicting resident values, integration with long-horizon planners, and combination with human-preference learning to personalize value weights. Taken together, these directions position RobotValues as an evaluation substrate for value-sensitive planning rather than as a complete solution to the problem of household robot alignment.