Benchmark RadarEgoGapBench
EgoGapBench evaluates egocentric action selection in multi-agent scenes, isolating the ability to choose actions from the agent's perspective when other agents are present. The benchmark includes training and test splits with human performance as reference.
- Released
- 2026-07-01
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Existing egocentric benchmarks conflate first-person view processing with perspective-taking, making it hard to isolate perspective understanding. EgoGapBench fills this gap by providing a controlled evaluation for a capability that is crucial for embodied AI and human-robot interaction, showing that state-of-the-art models fail at this task.
Motivation
Existing egocentric benchmarks have primarily constructed the egocentric setting from first-person-view data, which makes it difficult to evaluate egocentric perspective itself in isolation.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.