WAVE-0 Lets Robots Predict Consequences Before Acting

Introduction

Vision-language-action (VLA) policies have made robotic manipulation more general by combining visual perception, language understanding, and action generation. Yet many of these systems still select actions without explicitly representing their likely consequences. World action models (WAMs) add future prediction, but existing approaches often depend on expensive video generation or large-scale video pretraining. WAVE-0 proposes a lighter alternative: predict consequences in latent space and feed those predictions directly back into action generation.

Key points

• A bridge between VLAs and world models. Built on a pretrained VLA, WAVE-0 passes latent action plans generated by the vision-language model to a world model. The resulting future representations then condition the action expert, linking semantic planning, consequence prediction, and execution.
• No pixel-level video reconstruction. The framework uses a 257-million-parameter world model and does not require pixel reconstruction or large-scale video pretraining. Instead, it learns around real robot actions and their expected latent outcomes.
• Counterfactual ranking for training. Rather than reconstructing a complete future video, the world model is trained to rank alternative action consequences. This gives the policy a signal about which plans are more likely to lead to useful outcomes.
• Strong reported benchmark results. WAVE-0 reaches a 91.47% success rate on RoboTwin 2.0 and 98.7% on LIBERO. Its reported scores on LIBERO-Plus, LIBERO-Pro, and LIBERO-Para are 88.0%, 61.4%, and 75.0%, respectively. On real-world tasks, it raises the success rate of π₀.₅ from 38.3% to 73.3%.

Why it matters

The important design choice is that future prediction is not treated merely as an auxiliary training target. Predicted future representations directly influence the action expert at execution time. In effect, WAVE-0 gives a VLA policy a compact mechanism for asking, “If I take this action, what state is likely to follow?” before committing to the action.

The approach also suggests a practical route toward more predictive embodied policies. A useful world model may not need to begin with photorealistic video generation if a pretrained VLA already provides rich semantic representations and the training signal focuses on action-level alternatives.

The available material does not establish how the method will behave on longer-horizon tasks, major environmental changes, or different robot platforms. Still, the reported results indicate that latent future prediction can improve manipulation without imposing the full cost of video-based world modeling.

Source: Hugging Face Daily Papers