IT之家 · 智能时代T2·78 pts
- Proposes a 'perception-first' route, training on pure perception data without joint-action labels across 11 capabilities.
- Ranks first in comprehensive evaluation of 34 public perception benchmarks, scoring higher (73.68) than GPT-6 Astra (71.54).
- Aims to enable precise object localization (Grounding) to improve robot generalization and significantly reduce human demonstration costs.