Research map / Vision-language and multimodal / Robot action and navigation

World–Value–Action Model: Implicit Planning for Vision–Language–Action Systems

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes WAV, a VLA model that jointly learns a video world model, a trajectory value module, and an action decoder (all DiT-based, flow matching) and performs implicit planning by iteratively refining Gaussian noise distributions over video and value latents at inference, reaching 98.1% average on LIBERO.

Problem

Direct action prediction in VLA models does not reason over long-horizon trajectories or evaluate their consequences, limiting performance on complex, long-horizon manipulation.

Approach

Treats action generation as inference in a latent space of predicted future video and trajectory values: a world model rolls out futures, a value module scores them, and MPPI-style elite selection iteratively concentrates the latent noise distributions on high-value, feasible trajectories. A theoretical lemma argues action-space sampling has exponentially decaying feasible mass in horizon.

Key results

• WAV achieves 98.1% average success on LIBERO, vs 97.3% for VLA-Adapter and 96.4% for WAV without latent trajectory planning.
• On real-world Piper tasks (15 trials each), average success rises from 35.6% (GE-ACT) to 75.6% (WAV).
• Removing latent trajectory planning drops LIBERO-Long from 94.4 to 91.8, and average by 1.7 points.

Surprising findings

• On LIBERO-Long, WAV (94.4) does not beat VLA-Adapter (95.0) or UniVLA (94.0 tie-level), despite leading on average.
• The theoretical lemma and proposition are conditional and loosely stated; the proof of the key exponential-decay bound assumes intrinsic dimension scaling that is not justified empirically.

Limitations

• Time and storage overhead for deployment is the main limitation acknowledged by authors.
• Real-world evaluation is limited to three tasks on one dual-arm platform with 15 trials each.
• Theoretical results are conditional and not tight; the paper says it is not an optimality guarantee.

Experimental setup

Datasets: LIBERO real-world Piper demonstrations (drawer, bowl organization, towel flattening)

Benchmarks: LIBERO (Spatial, Object, Goal, Long) real-world Piper tasks

Metrics: success rate average success rate inference time GPU memory

Baselines: GE-ACT Diffusion Policy Octo OpenVLA SpatialVLA DiT Policy pi0-FAST GR00T-N1 pi0 VLA-0 Discrete Diffusion VLA MemoryVLA

Builds on: GE-ACT ReinboT MPPI T5-XXL Diffusion Transformer

Closest papers by meaning

• 3D-VLA: A 3D Vision-Language-Action Generative World Model (ICML 2024 · similarity 0.90)
• DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge (NeurIPS 2025 · similarity 0.90)
• WorldVLN: Autoregressive World Action Model for Aerial Vision-Language Navigation (NeurIPS 2026 · similarity 0.90)
• VLA-OS: Structuring and Dissecting Planning Representations and Paradigms in Vision-Language-Action Models (NeurIPS 2025 · similarity 0.90)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
• Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)