Summary
ATI-VLA tackles the common finding that predictive VLA models (ones that also forecast future states) underperform pure action-prediction models, attributing the gap to modality misalignment between observations/actions and conflicting joint-optimization objectives. It aligns observation and action representations through a shared discrete codebook, then injects the resulting predictive latents into action decoding via a lightweight side-path, reporting 97.9% average success on LIBERO (first among compared methods) and 70% on long-horizon real-robot tasks on a Galaxea R1 Lite.
Key Contributions
- Diagnosis of why predictive VLAs often lag pure action models: modality misalignment and joint-optimization conflict.
- A two-stage “Actionable Alignment Then Adaptive Injection” recipe: shared discrete codebook alignment, followed by lightweight side-path injection of predictive latents into the action decoder.
- State-of-the-art reported LIBERO result among compared predictive-VLA methods, plus long-horizon real-robot validation.
Strengths
- Directly addresses a known weakness in the predictive-VLA sub-field rather than just adding another prediction loss.
- The lightweight side-path design means the predictive signal can help without requiring the action decoder to be jointly optimized against a conflicting objective from scratch.
- Real-robot long-horizon evaluation, not just simulation benchmarks.
Weaknesses
- Numbers come from the authors’ own tables and are not independently verified.
- The side-path/injection mechanism adds architectural complexity, and its sensitivity to codebook size/design isn’t detailed in the available material.
Open Questions
- How much of the gain comes from the alignment stage versus the injection mechanism — is one necessary without the other?
- Does the approach generalize to other predictive targets (e.g., depth, affordance maps) beyond whatever future-state representation was used?
Significance
A targeted fix to a recognized failure mode in predictive VLA models, relevant to the broader debate (also touched by “Where Predictive Supervision Goes…”) about how best to fuse future-prediction objectives into action-generating policies.