Hasty Briefsbeta

Bilingual

GEN-1: Scaling Embodied Foundation Models to Mastery

3 days ago
  • #artificial-intelligence
  • #robotics
  • #machine-learning
  • GEN-1 is a general-purpose AI model achieving 99% success rates on simple physical tasks, up from 64% with previous models.
  • It completes tasks about 3 times faster than state-of-the-art and requires only 1 hour of robot data per result.
  • The model demonstrates mastery through reliability, speed, and improvisational intelligence, enabling recovery in unexpected scenarios.
  • Trained on over half a million hours of real-world data, GEN-1 shows commercial viability across a broad range of applications.
  • It builds on GEN-0's scaling laws for robotics, using human activity data from wearable devices instead of costly teleoperation datasets.
  • GEN-1 includes innovations in pre-training, post-training, reinforcement learning, multimodal guidance, and real-time inference techniques.
  • While not solving all tasks, it represents a step towards physical AGI, with improvements expected in future model generations.