SLIP-VLA: Single-Step Latent Imagination for Policy Learning in Vision-Language-Action Models

2026-09-27

AI Summary

This study addresses the inherent tension in Vision-Language-Action (VLA) models between the prohibitive computational cost of dense future prediction and the insufficient performance of single-step forecasting. To resolve this, we propose an efficient action prediction framework based on single-step latent imagination. Methodologically, a single-step denoising diffusion process generates dense future latents, while geometric-semantic alignment and an inverse dynamics module are incorporated to enhance representational sufficiency. Furthermore, an action-conditioned world model is introduced to optimize prediction quality. The proposed approach achieves state-of-the-art performance across both simulated and real-world robotic tasks with an inference latency of merely 12 milliseconds, effectively balancing computational efficiency with control precision.

Abstract

Vision-Language-Action models are increasingly effective for robotic manipulation, yet most predict actions directly from current observations without explicitly modeling future scene evolution. Recent methods introduce future prediction to improve action generation, but dense future modeling often requires expensive iterative denoising, while one-step alternatives can underperform their multi-step counterparts. To reconcile efficient future modeling with strong action performance, we present SLIP-VLA, a policy learning framework that equips VLA models with a Single-Step Latent Imagination for future-aware action prediction. SLIP-VLA obtains temporally dense future latent representations with a single denoising update, and we improve the perceptual sufficiency of these representations by aligning intermediate latents with future geometric and semantic features. We further improve their control sufficiency through action-conditioned latent world modeling and inverse dynamics modeling, explicitly coupling latent transitions with robot actions. SLIP-VLA achieves state-of-the-art performance across diverse simulation benchmarks and real-world manipulation tasks, while its single-step latent imagination takes only 12 ms.

Problem

• Vision-Language-Action models
• future prediction
• robotic manipulation
• latent imagination
• efficient modeling

Innovation

• Vision-Language-Action Models
• Single-Step Latent Imagination
• Latent World Modeling
• Inverse Dynamics Modeling
• Robotic Manipulation