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

ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection

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 ATI-VLA, which aligns observation and action latents in a shared VQ codebook and then injects frozen predictive observation latents into an OpenVLA-OFT backbone via an interval-based cross-attention side-path trained with only the action loss.

Problem

Predictive VLA models that forecast future observations or world knowledge often underperform direct action-prediction VLAs, due to a modality gap between observation and action representations and optimization conflicts from joint prediction and action losses.

Approach

Map observation and action into a shared discrete codebook via a unified predictive decoder (Stage 1), freeze it, then inject the observation latents into LLM action-query tokens every s layers through a lightweight cross-attention side-path trained only with the action loss (Stage 2).

Key results

• ATI-VLA reaches 97.9% average success on LIBERO, ranked first versus baselines such as F1 (95.7%) and UniVLA (95.5%).
• On RoboTwin 2.0 (four tasks), ATI-VLA averages 72.3% versus 60.0% for pi0 and 58.8% for OpenVLA-OFT (reproduced).
• On Galaxea R1 Lite real-world long-horizon tasks, ATI-VLA averages 70% versus 57% for F1 and 55% for OpenVLA-OFT; training cost is 16.2h versus 30.2h for OpenVLA-OFT.

Surprising findings

• Direct future prediction and injection without alignment both yield limited or degraded performance relative to the OpenVLA-OFT-style baseline in the controlled same-backbone study.
• Dense injection at every LLM layer hurts long-horizon performance (s=1 gives 92.8% on Long versus 96.2% at s=8).

Limitations

• Real-world evaluation covers only five long-horizon tasks on two platforms with 25 trials per stage.
• Inference does not produce explicit future predictions, and the alignment module must be retained at test time.
• Injection interval and optimal layer depend on backbone and are tuned empirically.

Experimental setup

Datasets: LIBERO RoboTwin 2.0

Benchmarks: LIBERO RoboTwin 2.0 Galaxea R1 Lite real-world tasks AgileX Cobot Magic real-world tasks

Metrics: success rate stage-wise success rate rank latency throughput training cost (h) cross-modal similarity CKNNA

Baselines: OpenVLA-OFT pi0 PD-VLA ATM Seer DreamVLA CoT-VLA F1 H-GAR UniVLA WorldVLA CogVLA

Builds on: OpenVLA-OFT OpenVLA VQ-VAE

Closest papers by meaning

• ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.92)
• Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
• VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.91)
• RotVLA: Rotational Latent Action for Vision-Language-Action Model (NeurIPS 2026 · similarity 0.91)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
• NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)