THAW-VLA: World-Model Features Without World-Model Latency
from Embodied AI 101 Hosted by Shaoqing Tan
This work distills frozen world-model representations into compact VLAs via a single feature-alignment loss at training time, with the teacher cached once and absent at inference; a 0.8B student achieves 97.9% on LIBERO and lifts RoboCasa-GR1 humanoid performance from 48.2% to 50.5% at zero extra runtime cost (32 ms / 1.86 GB on RTX 5090). The approach generalizes across scales and architectures on both single-arm and bimanual real hardware. In this episode of Embodied AI 101, we explore "THAW-VLA: World-Model Features Without World-Model Latency". We break down the research, methodology, and real-world implications for robotics, AI, and physical intelligence. Embodied AI 101 covers the latest research at the intersection of AI and physical intelligence — robotics, manipulation, world models, and the path from digital intelligence to embodied agents.
Embed this episode
What this episode covers
This work distills frozen world-model representations into compact VLAs via a single feature-alignment loss at training time, with the teacher cached once and absent at inference; a 0.8B student achieves 97.9% on LIBERO and lifts RoboCasa-GR1 humanoid performance from 48.2% to 50.5% at zero extra runtime cost (32 ms / 1.86 GB on RTX 5090). The approach generalizes across scales and architectures on both single-arm and bimanual real hardware.
No transcript yet.