VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding

TL;DR

VFlash accelerates visual planning in AI systems by decoding geometric primitives like boxes and trajectories in a single forward pass, leveraging their structural properties for faster and more accurate predictions. It bypasses sequential token generation used in autoregressive models, enabling significant speedups without sacrificing spatial precision. The method improves performance on robot trajectory tasks and enhances success rates in world action models.

Why It Matters

Faster and more accurate visual planning enables real-time decision-making in robotics and autonomous systems, where delays or inaccuracies can compromise safety and efficiency. By reducing computational overhead while maintaining high precision, VFlash makes complex spatial reasoning more feasible for practical deployment.

Key Findings

• VFlash achieves faster generation by predicting entire visual primitives in one forward pass using a lightweight drafter that respects primitive structure.
• The method reduces prediction errors compared to autoregressive decoding across 2D and 3D trajectory tasks, without requiring exact token-level alignment with ground truth.
• Single-rollout evaluations show approximately 25x speedup for 3D point motion forecasting, and a world action model using VFlash plans reaches 70.2% average success across benchmarks.