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

Yijie Zhu; Rui Shao; Jie He; Wei Li; Bo Zhao; Yelin Wang; Xiaochen Yuan; Tao Tan

cs.CVOct 1, 2026

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

Authors: Yijie Zhu, Rui Shao, Jie He, Wei Li, Bo Zhao, Yelin Wang, Xiaochen Yuan, Tao Tan, +3 more

Organizations: Harbin Institute of Technology, Shenzhen · Institute for Artificial Intelligence, Great Bay University · Shenzhen Loop Area Institute · Macao Polytechnic University · Shenzhen Technology University · Dongguan Key Laboratory for Intelligence and Information Technology

Abstract

Predictive Vision-Language-Action (VLA) models aim to improve robotic manipulation via future observation or world dynamics forecasting. However, existing approaches often fail to realize this potential and underperform direct action prediction models. We argue that these limitations stem from modality misalignment between observations and actions, together with joint optimization conflicts that drive learning away from an action-centric objective. To this end, we introduce ATI-VLA, an Action-Centric Predictive Vision-Language-Action framework via Actionable Alignment Then Adaptive Injection. Specifically, it follows a two-step design: 1) Actionable Representation Alignment via a Shared Codebook. It aligns predictive observation and action representations by mapping both modalities into a shared discrete latent space via a unified codebook, making predictive observation latents readily usable for action generation and mitigating modality misalignment. 2) Action-Centric Adaptive Injection of Predictive Latents. Building upon this, it then injects predictive observation latents into action decoding as explicit predictive priors via a lightweight adaptive side-path, enabling adaptive predictive guidance under a single action-centric objective. Extensive experiments on both simulation and real-world robotic tasks demonstrate that ATI-VLA achieves state-of-the-art performance with faster convergence.

Figures & tables

Figure 1 : Overview of VLA paradigms and motivation for ATI-VLA . Figures (a)–(c) compare general (non-predictive) VLA, predictive VLA, and our proposed action-centric predictive framework, ATI-VLA . Figures (d)–(f) identify key limitations of existing predictive VLA approaches, including observation–action modality gaps and slow convergence induced by joint optimization, which are responsible for their inferior performance relative to general VLA. By addressing these limitations, ATI-VLA enables the predictive VLA paradigm to more fully realize its potential, achieving state-of-the-art performance with faster convergence.

Figure 2 : Step 1: Actionable Representation Alignment. Left: The current observation and action are encoded and vector-quantized into a shared codebook, then decoded by a unified attention-based decoder to jointly predict the future observation $\mathcal{O}_{t+n}$ and action $\mathcal{A}_{t+n}$ . Right: t -SNE visualization of observation (main/wrist) and action latents during training: they are initially dispersed and modality-separated, but progressively converge into an aligned shared structure after training.

Figure 3 : Step 2: Action-Centric Adaptive Injection. Predictive observation latents are adaptively transformed via a lightweight side-path and injected into LLM layers to modulate action query tokens. Interval-based injection provides effective predictive guidance while preserving stable, action-centric decoding.

Category	Method	Success Rate( $\uparrow$ )					Rank( $\downarrow$ )
Category	Method	Spatial	Object	Goal	Long	Average	Rank( $\downarrow$ )
General VLA	OpenVLA-OFT [RSS’25] [ 20 ]	97.6%	98.4%	97.9%	94.5%	97.1%	2
	$\pi_{0}$ [RSS’25] [ 2 ]	96.8%	98.8%	95.8%	85.2%	94.2%	7
	PD-VLA [IROS’25] † [ 44 ]	95.5%	96.7%	94.9%	91.7%	94.7%	6
Predictive VLA	ATM [RSS’24] [ 50 ]	68.5%	68.0%	77.8%	39.3%	63.4%	11
	Seer [ICLR’25] [ 45 ]	-	-	-	87.7%	-	-

Table 1 : Simulation Results on LIBERO. Comparison of task success rates and their ranks.

Table 5

Galaxea R1 Lite platform
Method	Object P		Markers C		Plate H			T-shirt F			T-shirt F(OOD)			Average
Method	Open	+Pack	C1	+C2	Pick	+Pass	+Place	S1	+S2	+S3	S1	+S2	+S3	Average
OpenVLA-OFT†	18/25	17/25	17/25	15/25	18/25	17/25	15/25	17/25	15/25	14/25	10/25	8/25	8/25	55%
UniVLA†	14/25	13/25	13/25	13/25	15/25	14/25	12/25	15/25	15/25	12/25	13/25	10/25	7/25	46%
$\mathcal{F}_{1}$ †	16/25	15/25	19/25	17/25	15/25	14/25	14/25	17/25	15/25	14/25	16/25	12/25	11/25	57%
DreamVLA†	15/25	14/25	17/25	14/25	13/25	11/25	11/25	15/25	14/25	14/25	13/25	12/25	10/25	50%

Table 4 : Real-World Experiments on Different Platforms. The table reports success rates at each stage for five long-horizon tasks, including an out-of-distribution (OOD) variant of T-shirt Folding. Results marked with “†” were reproduced using the same experimental settings as ATI-VLA. The average success rate corresponds to the completion rate of the entire task sequence.

Table 7

Figure 4 : Ablation on Actionable Representation Alignment Strategy. t-SNE of observation/action latents under different alignment strategies: (a) dual decoder, (b) sequential decoding, (c) dual codebook, (d) our method (shared codebook + joint decoding).

Panel of Figure 5 — Figure 5 : Visualization results of the proposed method. (a) shows future observation prediction results. (b) illustrates step-by-step execution of real-world manipulation tasks.

Table 10

Figure 6 : Convergence Comparison. ATI-VLA converges faster and performs better than ATI-VLA w/o alignment, while Fig. 1(f) further shows faster convergence over prior predictive VLA baselines.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Parameters	FLOPs	LIBERO	RoboTwin 2.0	Real World	Cost (h)
OpenVLA-OFT (Baseline)†	7.71 B	8.45 T	96.6	58.8	55.0	30.2
ATI-VLA	8.04 B (+4.2%)	8.51 T (+0.7%)	97.9 (+1.3)	72.3 (+13.5)	70.0 (+15.0)	16.2 (-46.4%)

Appendix

Table 10 : Trade-off Between Model Complexity and Task Performance.

Method	LIBERO	RoboTwin 2.0	Real-World
Baseline + Direct Future Prediction	95.8%	52.4%	48%
Baseline + injection without alignment	93.4%	51.3%	50%
Baseline + with alignment but without adaptive injection	96.2%	60.7%	62%
ATI-VLA (Ours)	97.9%	72.3%	70%

Appendix

Table 11 : Controlled Same-Backbone Ablation. All variants use the same backbone, the same training data, and the same evaluation protocol.

Method	LIBERO	RoboTwin 2.0	Real-World
Image Prediction (CoT-VLA-like)	95.2%	52.0%	50%
World-Knowledge Prediction (DreamVLA-like)	94.4%	50.3%	46%
ATI-VLA (Ours)	97.9%	72.3%	70%

Appendix

Table 12 : Controlled Comparison over Predicted Modalities. All variants use the same backbone, training data, and evaluation protocol, differing only in the predicted modality and predictive design. Results report average success rates.

LLM	Architectural Prior	Norm	Layers	Best $s$
LLaMA-2-7B	Dense causal attention	RMSNorm	32	8 $(1/4)$
Mistral-7B	Sliding-window attention	RMSNorm	32	8 $(1/4)$
OPT-1.3B	Dense causal attention	LayerNorm	24	8 $(1/3)$

Appendix

Table 13 : Generalization of Injection Interval Across Backbones. We evaluate LLM backbones with different attention patterns and normalization schemes. The preferred injection interval remains around one-third to one-fourth of the model depth.

Setting	LIBERO	RoboTwin 2.0	Real-World
Effect of Action Information in Stage 1
Observation-only Alignment	95.6%	64.3%	61%
Full Alignment (Ours)	97.9%	72.3%	70%
Effect of Predictive Alignment Objective
Reconstruction-based Alignment	96.0%	62.4%	59%
Predictive Alignment (Ours)	97.9%	72.3%	70%

Appendix

Table 14 : Ablation on Stage-1 Alignment Design. We examine whether action information and predictive learning are necessary for learning action-grounded latent representations.

Method	Semantic	Pose/Control	Transition	Success
w/o Alignment	26	33	49	92
ATI-VLA (Ours)	11	19	41	129

Appendix

Table 15 : Failure-Mode Diagnosis in Long-Horizon Real-World Tasks. We categorize failures on Plate Handover and T-shirt Folding, with 100 trials per task and 200 trials in total for each method. The non-aligned variant fails more often in semantic grounding, pose/control grounding, and transition-phase stages.

Method	Safe Steps $\uparrow$	Pre-Failure Steps $\uparrow$	Gap $\downarrow$
w/o Alignment	0.59	0.37	0.22
ATI-VLA (Ours)	0.80	0.70	0.10

Appendix

Table 17 : Cross-Modal Similarity Before Failure. We compare cross-modal similarity on safe steps and pre-failure steps. The non-aligned variant exhibits a larger alignment drop before failure, while ATI-VLA maintains higher alignment in these high-risk regions.

Method	$\textbf{Avg. Cos}(g_{\rm pred},g_{\rm act})\uparrow$	Neg. Ratio $\downarrow$	RoboTwin 2.0 $\uparrow$	Real-World $\uparrow$
DreamVLA	-0.23	43%	–	50%
Direct Future Prediction	-0.29	39%	52.4%	48%
Direct Future Prediction + PCGrad	-0.09	21%	57.6%	55%
ATI-VLA (Align + Inject)	Avoided by design	Avoided by design	72.3%	70%

Appendix

Table 18 : Gradient Conflict Between Prediction and Action Objectives. We measure the gradient cosine similarity between prediction and action losses on shared parameters. PCGrad partially alleviates the conflict, while ATI-VLA avoids it by design and achieves the best performance.

Figure 7 : Extended Future Observation Prediction from Aligned Action Representations. Additional qualitative examples of future observation prediction on the LIBERO and RoboTwin 2.0 benchmarks. Each triplet shows the current observation, ground-truth future observation, and the predicted future observation from the aligned action representation.

Figure 8 : Extended Step-by-Step Real-World Task Execution. Further visualization of all real-world tasks. Each frame captures a critical stage in the sequential execution.

Explore similar work

LARA: Latent Action Representation Alignment for Vision-Language-Action Models
Jun 5, 2026Mengya Liu, Baoxiong Jia, Jiangyong Huang +2Latent Action Models Large-Scale Robot Demonstration Datasets
State Key Laboratory of General Artificial Intelligence, BIGAI · Peking University · Delta Intelligence
VOTE: Vision-Language-Action Optimization with Trajectory Ensemble Voting
Jul 7, 2025Juyi Lin, Amir Taherin, Arash Akbari +11Diffusion-Based Vision-Language-Actions Robotic Manipulation
Northeastern University, Boston, USA · EmbodyX,San Mateo,USA
GuidedVLA: Specifying Task-Relevant Factors via Plug-and-Play Action Attention Specialization
May 12, 2026Xiaosong Jia, Bowen Yang, Zuhao Ge +17Recent Vision-Language Models Guidance
Institute of Trustworthy Embodied AI (TEAI), Fudan University · 2Shanghai Key Laboratory of Multimodal Embodied AI · Shanghai Jiao Tong University +1

Jun 5, 2026, cs.CVJ/K move · Enter open · S save

LARA: Latent Action Representation Alignment for Vision-Language-Action Models

Mengya Liu, Baoxiong Jia, Jiangyong Huang, Jingze Zhang+1

State Key Laboratory of General Artificial Intelligence, BIGAI · Peking University · Delta Intelligence

Latent Action Models Large-Scale Robot Demonstration Datasets

Open paper AskPDF arXiv

Visual-language action (VLA) models enable robots to predict actions directly from observations and language instructions, but their performance depends on large-scale, high-quality data and is limited by the scarcity of real-world robot action datasets. To facilitate VLA model learning with abundant unlabeled human videos, Latent Action Models (LAM) learn latent action representations from visual dynamics to provide additional supervision for VLA learning. However, LAM and VLA are typically trained separately, leaving LAM ungrounded during VLA training and VLA models constrained by frozen LAM representations. To address these issues, we propose Latent Action Representation Alignment (LARA), a plug-and-play framework that jointly optimizes LAM and VLA via representation alignment. This enables reciprocal benefits where LAMs learn with action trajectories to avoid spurious visual changes, while VLAs are regularized by forward dynamics learned within LAMs to reduce hallucinations of functionally ineffective trajectories. We demonstrate LARA versatility and effectiveness for pre-training, post-training enhancement of pre-trained VLA models, and LAM refinement, achieving an average of ~10%, ~5%, and ~15% improvement over 3 simulation and 1 meticulously designed real-world robotic manipulation benchmarks.

Abstract

Explore similar work

LARA: Latent Action Representation Alignment for Vision-Language-Action Models

VOTE: Vision-Language-Action Optimization with Trajectory Ensemble Voting

GuidedVLA: Specifying Task-Relevant Factors via Plug-and-Play Action Attention Specialization