X-Foresight: A Joint Vision-Action Causal Forecasting Network via Predictive World Modeling
Researchers have developed X-Foresight, a new predictive world model integrated into Vision-Language-Action (VLA) models. This model aims to equip VLA systems with physical world knowledge by predicting future video sequences, addressing challenges of trivial extrapolation and long-term causality. X-Foresight utilizes a chunk-wise auto-regressive strategy and temporal importance sampling to learn world dynamics and causality more effectively, outperforming existing VLA baselines in planning tasks. AI
IMPACT Enhances VLA models with physical world knowledge, potentially improving autonomous system planning and generalization.