(논문 요약) π0: A Vision-Language-Action Flow Model for General Robot Control (Paper)

핵심 내용

• Architecture: pretrained PaliGemma vision (3B) + action expert (300M)
  • $q_t$: vector of joint angles

• 학습: flowing matching

• $v_{\theta}$: network
• $A^{\tau}_t$: noisy action
• $u(A^{\tau}_t|A_t)$: denoising vector field

실험 결과