Benchmark RadarEgo-METAS
Ego-METAS evaluates online temporal action segmentation in egocentric video, where models must select sensor modalities (RGB, audio, gaze, IMU, monochrome) per timestep to maximize accuracy under hardware-representative energy budgets. It includes 100+ hours of untrimmed video from multiple datasets.
- Released
- 2026-05-29
- Readiness
- Inspectable
- Primary field
- Robotics & Autonomous Systems
Why it matters
This benchmark addresses the gap in energy-aware perception for embodied AI by providing a standardized testbed for developing and comparing cost-aware sensor routing policies in continuous, untrimmed environments. It enables assessment of trade-offs between predictive accuracy and energy consumption, with practical implications for always-on devices.
Motivation
To operate in the physical world, embodied agents must perceive their environment in an "always-on" fashion, selectively accessing the most informative sensors to balance energy constraints and task accuracy.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.