REAL: A Unified Evaluation
of Physical AI
Abstract
Physical AI evaluation requires a system that relates capability definitions to observable task evidence. We present REAL, a unified capability–task–metric framework organized around basic intelligence, adaptive learning, human–machine coexistence, safety and reliability, and application practice. The framework organizes evaluation design through 14 dimensions, 55 abilities, and 158 capability indicators. Its five-level hierarchy of domains, dimensions, abilities, assessment points, and observations guides their operationalization as task specifications. Four leaderboards instantiate the system through embodied reasoning, simulated manipulation, physical manipulation, and world prediction. Theoretical applicability, scheduled scope, and completed experiments are reported separately. Shared task identities, frozen protocols, auditable artifacts, coverage, and uncertainty support reproducible comparisons, with dedicated simulation-to-real alignment and model-role studies. The report describes the complete capability system and its operational mapping; empirical findings require verified model submissions and measured results.
- Basic
intelligence - Adaptive
learning - Human–machine
coexistence - Safety and
reliability - Application
practice
Domains → dimensions → abilities → assessment points → observations
Tasks / Metrics / Benchmarks / Auditable evidence