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

1 Introduction

• Existing predictive VLA approaches generally underperform non-predictive baselines; modality gaps and joint-training conflicts weaken action-centric learning and slow convergence.Competing prediction and action gradients can divert optimization from the action objective and, in extreme cases, cause training collapse.Source: §1 · pp. 1–2
• ATI-VLA first aligns predictive observations and actions through a shared codebook, then injects predictive latents into action decoding through a lightweight adaptive side-path.Injection is optimized solely with the action loss, using predictive information as guidance rather than a competing objective.Source: §1 · p. 2
• Across simulation and real-world tasks, ATI-VLA achieves state-of-the-art performance and faster convergence than existing predictive approaches.The introduction reports this as an overall result, without providing numerical comparisons here.Source: §1 · pp. 2–3
• The framework’s contributions include an action-centric predictive design and a shared-codebook alignment that grounds predictive observations in the action space.The adaptive injection strategy integrates predictive information into decoding under a single action-centric objective.Source: §1 · p. 3

2.1 Preliminary: Predictive VLA

• Predictive VLA models use future observation or world-dynamics forecasts to guide action generation, typically optimizing action and prediction jointly.The paper identifies modality gaps and competing objectives as challenges to preserving action-centric learning.Source: §2.1 · p. 3

2.2.1 OVERVIEW

• ATI-VLA trains alignment first, then freezes it while training a lightweight injection side-path using only the action objective.The alignment stage learns a unified actionable latent space; the subsequent injection stage adds predictive latents to action decoding.Source: §2.2.1 · p. 3, §2.2.3 · p. 6

2.2.2 Actionable Representation Alignment

• A shared codebook maps predictive observation and action representations into a discrete latent space, and a unified decoder jointly predicts future observations and actions.The alignment objective combines observation- and action-prediction losses; training latents progressively co-locate from initially modality-separated distributions.Source: §2.2.2 · pp. 4–5

2.2.3 Action-Centric Adaptive Injection

• A lightweight side-path adaptively injects predictive latents into LLM action-query states at regular layer intervals, rather than every layer.At injection layers, action-token representations query predictive latents, and a learnable scalar controls guidance strength; dense injection can over-constrain decoding.Source: §2.2.3 · pp. 5–6

3.1 Experimental Setup

• Experiments cover LIBERO and RoboTwin 2.0 simulations and real-world long-horizon tasks, using generalist and predictive VLA baselines.Real-world evaluation includes the Galaxea R1 Lite and AgileX Cobot Magic platforms, five tasks across four categories, and an out-of-distribution T-shirt-folding variant.Source: §3.1 · pp. 6–7

3.2 Overall Performance

• ATI-VLA achieves the best overall rank on LIBERO and the highest average success rate on RoboTwin 2.0, while reaching 70% average success on Galaxea R1 Lite long-horizon tasks.On LIBERO, the Object and Long suites reach 99.4% and 96.2%; on Galaxea it exceeds DreamVLA (50%), F1 (57%), and OpenVLA-OFT (55%).Source: §3.2 · p. 7, Table 2 · p. 7, Table 4 · p. 7
• 16.2h is ATI-VLA’s total training cost for both stages, the lowest among the compared methods; its inference efficiency is only marginally lower than OpenVLA-OFT.The reported training times are 30.2h for OpenVLA-OFT, 33.5h for DreamVLA, and 67.4h for CoT-VLA; removing the side-path adds negligible overhead.Source: §3.2 · pp. 7–8, Table 3 · p. 7

3.3 Ablation Studies

• Only the shared-codebook, joint-decoding alignment produces a well-mixed observation–action latent distribution, with improved cross-modal similarity and CKNNA.The comparison includes shared-codebook dual decoders, shared-codebook sequential decoding, and dual codebooks with joint decoding.Source: §3.3 · p. 8, Figure 4 · p. 8, Table 7 · p. 8
• Random latent injection substantially degrades performance, while injection every eight layers balances predictive guidance and action-decoding autonomy.Dense injection at s = 1 hurts long-horizon performance, whereas sparse injection at s = 16 weakens predictive conditioning; the preferred interval remains near one-third to one-fourth of model depth across backbones.Source: §3.3 · p. 8, Table 5 · p. 8, Table 6 · p. 8
• Controlled comparisons attribute the gains mainly to the full align-then-inject design rather than backbone, training-data, evaluation-protocol, or predicted-modality differences.The same-backbone comparison includes direct future prediction, injection without alignment, alignment without adaptive injection, and full ATI-VLA.Source: §3.3 · pp. 8–9, Table 8 · p. 9

3.4 Qualitative Analysis

• Predicted frames show structural fidelity and temporal coherence, while real-world sequences illustrate multi-stage manipulation and convergence curves show faster optimization and higher success than the unaligned variant.Examples include bag opening, inter-arm plate transfer, and precise placement; ATI-VLA also has faster optimization and better final performance than prior general and predictive VLA baselines.Source: §3.4 · p. 9, Figure 5 · p. 9

Vision–Language–Action Models

• The broader VLA line developed from RT-2 [4] integrating pretrained VLMs, through RT-X scaling with large-scale data, to later models including OpenVLA [21], OpenVLA-OFT [20], and the π-series [18].Figure 1(f) reports faster convergence for ATI-VLA over prior predictive VLA baselines.Source: §4 · pp. 9–10, Figure 6 · p. 10

Predictive VLA Models

• Predictive VLA approaches add future prediction to VLA models, but can suffer from modality misalignment and joint-optimization conflicts that weaken action-centric learning.ATI-VLA instead aligns predictive latents with actions and injects them under a single action-centric objective.Source: §4 · p. 10

Robotic Manipulation in Learned Latent Spaces

• ATI-VLA’s shared codebook jointly predicts future observations and actions to learn shared action-grounded prototypes, rather than reconstructing inputs or using observation latents as output action tokens.Unlike VITA [36], ATI-VLA uses a unified decoder, an adaptive side-path, and a frozen alignment module with only the action loss in Stage 2.Source: §4 · p. 10

5 Conclusion

• ATI-VLA first aligns observations and actions in a shared discrete latent space, then adaptively injects predictive latents under a single action objective, achieving state-of-the-art performance with faster convergence.The conclusion presents this as an Actionable Alignment Then Adaptive Injection framework for predictive VLA.Source: §5 · p. 10

Appendix

• The appendix covers implementation and training, benchmarks and evaluation, extended ablations and analyses, theoretical analyses, limitations, future directions, and societal implications including deployment risks.Its sections also include failure modes and additional qualitative results.Source: Appendix · p. 15

A.1 Model Details

• The alignment stage quantizes observation and action representations through one codebook, then a unified decoder jointly predicts future observations and actions.The observation encoder processes a single frame; the action encoder projects the current action vector into a spatially aligned embedding grid.Source: §A.1 · p. 15
• At inference, ATI-VLA does not decode future observations explicitly; it maps the current observation into the learned latent space and injects that predictive latent into the action decoder.The alignment module is trained with future-observation and action supervision, then frozen and retained at inference time; the implementation uses OpenVLA-OFT with an OpenVLA-7B backbone.Source: §A.1 · p. 15

Simulation Training Setup

• LIBERO training uses LoRA with rank 32 and α = 64, batch size 64, learning rate 5 × 10^-4, and action sequence length n = 8; RoboTwin 2.0 uses batch size 32 and action chunks of 25.LIBERO validation occurs every 5,000 steps, and the highest-validation-performance checkpoint is selected.Source: §A.2 · p. 16

Real-World Training Setup

• Real-world training uses three camera views, LoRA with rank 32 and α = 64, action chunks of n = 25, and 80,000 steps with batch size 32.The learning rate decays from 5 × 10^-4 to 5 × 10^-5 after 50,000 steps; checkpoints are evaluated every 5,000 steps from step 60,000, and the best-performing checkpoint is retained.Source: §A.2 · p. 16

• LIBERO evaluates complementary robotic capabilities through multiple language-conditioned manipulation suites, each containing 10 tasks.The suites emphasize spatial reasoning, object-level generalization, goal-conditioned behavior diversity, and long-horizon temporal reasoning and sequential control.Source: §B.1 · p. 16
• Evaluation follows the standard LIBERO protocol, using 50 expert demonstrations per task from the official release.Source: §B.1 · p. 16
• RoboTwin 2.0 covers bimanual manipulation with 50 dual-arm tasks, while evaluation selects four tasks and runs 100 trials per task.Its benchmark includes 731 objects across 147 categories and structured domain randomization.Source: §B.1 · p. 16

B.2 Real-World Tasks

• ATI-VLA is evaluated on five long-horizon sequential tasks across four capability categories, including an out-of-distribution T-shirt-folding variant.Evaluation uses the Galaxea R1 Lite and AgileX Cobot Magic platforms, progressive stage success rates, and 50 human-teleoperated demonstrations per task.Source: §B.2 · pp. 16–17
• The tasks test bimanual coordination, precision manipulation, object transfer, and deformable-object handling through staged operations.They include Object Packing, Markers Collection, Plate Handover, and T-shirt Folding; the OOD variant changes shirt color and dimensions while retaining the same folding procedure.Source: §B.2 · p. 17
• Real-world training uses 50 expert demonstrations per task, with moderate augmentation for four tasks and fixed fabric type and layout for in-distribution T-shirt folding.Augmentation varies initial poses, spatial arrangements, and environmental configurations.Source: §B.2 · p. 17

B.3 Baselines

• Baselines span general VLA models that map current observations and language directly to actions, and predictive models that use look-ahead signals to guide action generation.Examples include OpenVLA-OFT and π0 for general VLA, and Seer, DreamVLA, CoT-VLA, F1, and H-GAR for predictive VLA.Source: §B.3 · pp. 17–18
• The comparisons also include CogVLA, PD-VLA, UniVLA, and an ATI-VLA variant that removes explicit actionable representation alignment.The ablated variant jointly models future observations and actions without the alignment stage.Source: §B.3 · p. 18

C.1 Extended Ablations

• 46.4% lower training time (30.2 to 16.2 hours) accompanies 0.7% additional FLOPs while ATI-VLA improves task performance over OpenVLA-OFT.The comparison is reported in Table 10.Source: §C.1 · p. 18
• In a controlled same-backbone comparison, direct future prediction yields limited improvement, injection without alignment degrades performance, and full ATI-VLA performs best across simulation and real-world settings.The variants share the backbone, training data, and evaluation protocol; the reported gains primarily stem from combining actionable alignment and adaptive injection.Source: §C.1 · pp. 18–19
• Image-prediction and world-knowledge-prediction variants both underperform ATI-VLA, especially on RoboTwin 2.0 and real-world tasks.The comparison suggests the gains do not simply come from predicting a particular modality.Source: §C.1 · p. 19, Table 12 · p. 19
• The preferred injection interval remains around one-third to one-fourth of model depth across tested backbones; dense injection can over-constrain decoding, while sparse injection can provide insufficient guidance.The exact optimum may vary slightly across backbone families.Source: §C.1 · p. 19, Table 13 · p. 19
• Alignment using both observations and actions outperforms observation-only alignment, and predictive alignment outperforms reconstruction-based alignment across settings.The study attributes the difference to action-side supervision and predictive alignment’s focus on action-induced state transitions.Source: §C.1 · p. 19, Table 14 · p. 20
• ATI-VLA achieves a 97.9% average success rate on LIBERO with 0.1% standard deviation across random seeds.The reported small variance across suites suggests the gains are not due to random fluctuations.Source: §C.1 · p. 19

C.2.1 Modality Gaps.

• Non-aligned failures include wrong target selection, inaccurate approach poses, mis-oriented placements, and failures to switch actions after subgoal transitions.First dominant failures cluster in later or transition-heavy stages, such as Plate Handover’s Pass/Place and later T-shirt-folding stages.Source: §C.2.1 · p. 20, Table 16 · p. 21
• Cross-modal similarity drops by 0.22 before failure for the non-aligned variant, compared with a 0.10 gap for ATI-VLA.The comparison uses safe steps and pre-failure steps and reports higher similarity for ATI-VLA in both regions.Source: §C.2.1 · p. 20, Table 17 · p. 21

C.2.2 Optimization Conflicts.

• Prediction and action losses show conflicting gradients; PCGrad shifts average cosine similarity from -0.29 to -0.09, while ATI-VLA’s second stage uses only the action loss.The reported conflict study finds negative average cosine similarity and a high ratio of negative gradient pairs for DreamVLA and direct future prediction.Source: §C.2.2 · pp. 20–21

C.2.3 Why Is the Modality Gap Severe in Predictive VLA Settings.

• Future-observation prediction can encode visually rich details irrelevant to control, while action generation needs control-relevant factors such as target identity, contact state, object pose, and subgoal transitions.Source: §C.2.3 · p. 21
• ATI-VLA uses observation and action information to learn an action-grounded shared latent space, then injects aligned latents under a single action objective.This design avoids further prediction–action optimization conflict and makes predictive information more directly usable for control.Source: §C.2.3 · p. 21

C.3 Additional Qualitative Results

• Across LIBERO and RoboTwin 2.0, predicted frames preserve scene structure, temporal coherence, object details, and spatial relations, with robustness across task suites suggesting transfer beyond training distribution.Source: §C.3 · pp. 21–22, Figure 7 · p. 22
• Table 18 reports that PCGrad partially alleviates prediction–action gradient conflict, while ATI-VLA avoids it by design and achieves the best performance.Source: Table 18 · p. 22
• Real-world sequences show dual-arm object packing and plate handover, precise marker insertion, and standard and OOD T-shirt folding, illustrating extended planning and bimanual synchronization.Source: §C.3 · pp. 22–23, Figure 8 · p. 23

D.1 Shared Discrete Codebook

• Under an L-Lipschitz action loss, excess risk from using predictive observation latents is upper-bounded by latent discrepancy, which the shared codebook is intended to reduce.Source: §D.1 · pp. 23–24

D.2 Interval Injection

• Interval injection trades predictive guidance against perturbation accumulation: dense injection adds more perturbation sources, while overly sparse injection provides too few signals.Source: §D.2 · p. 25
• The interval analysis is a local trend argument, not a tight global bound or exact optimizer; its preferred L/3–L/4 range is reported as empirical across backbones.Source: §D.2 · p. 25

E Limitations and Future Work

• Current training and evaluation cover a relatively limited set of manipulation datasets and task scenarios, with real-world experiments relying mainly on task-specific teleoperated demonstrations.Source: E Limitations and Future Work · p. 25

F Broader Impact and Potential Risks

• The paper identifies assistive robotics, warehouse automation, flexible manufacturing, and action-centric predictive learning in embodied AI as potential beneficiaries.Source: F Broader Impact and Potential Risks · p. 25
• Stronger robotic policies may pose risks in unconstrained or safety-critical settings if prediction errors, distribution shifts, or incorrect instruction grounding lead to unintended behavior.Source: F Broader Impact and Potential Risks · p. 25