VLA 深度追蹤VLA 深度追踪

The central paradigm war in embodied AI. VLA (Vision-Language-Action) maps observations directly to actions end-to-end — simple, scalable, but needs massive data and generalizes poorly. WAM (World-Action Model) first learns how the world works, then plans actions through mental simulation — better generalization and data efficiency, but world models are often inaccurate. The boundary is blurring: Pi0.5 uses flow matching (generative, WAM-like), GR00T adds video prediction. The winner likely is a hybrid.

Both generate continuous actions from the same VLA backbone but take different mathematical routes: diffusion iteratively denoises random noise into actions (slow, expressive), while flow matching uses optimal transport for a direct trajectory (fast, efficient). If flow matching matches diffusion quality, it could replace it as the default action head.

After pretraining a VLA, two competing strategies exist: SFT directly imitates expert demonstrations (simple, stable), while RL fine-tuning (GRPO/DPO) optimizes a reward signal to go beyond the demonstration distribution. RL can discover novel strategies but is harder to stabilize.

World models learn by predicting the future (imagination-based planning), while RL learns from reward feedback. If world models become accurate enough, they could reduce the need for expensive real-world RL exploration.

Two approaches to dexterous manipulation: tactile sensing adds explicit touch feedback (hardware cost, rich signal), while dexterous hand control relies on proprioception and vision alone (simpler hardware, harder control). The winner depends on sensor cost-to-performance ratio.

Sim-to-Real trains one robot in simulation then transfers (cheap data, reality gap risk), while Cross-Embodiment trains across multiple real robots directly (expensive data, natural generalization). The approaches represent different bets on where generalization should happen.