Benchmark RadarNormAct
NormAct evaluates embodied agents on 550 TongSim scenarios where the same goal permits norm-compliant or norm-violating action sequences, testing whether agents infer and apply scene-relevant social norms during ordinary tasks.
- Released
- 2026-06-26
- Readiness
- Paper only
- Primary field
- Robotics & Autonomous Systems
Why it matters
NormAct addresses the gap between goal achievement and proactive norm compliance, providing a way to assess whether embodied agents respect unstated social norms without prompting, which is critical for real-world deployment.
Motivation
Embodied agents driven by multimodal large language models (MLLMs) can often complete everyday tasks from visual observations, but goal achievement does not establish whether they proactively respect unstated social norms.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.