Hyperspectral imaging can distinguish objects that look similar in ordinary RGB, but a pretrained vision-language-action (VLA) model's frozen decoder is not built to consume raw hyperspectral data. SPECTRA-VLA addresses this by injecting spectral information into the token stream without retraining the visual encoder or action decoder.

The method adds a learned continuous residual to each target token, guided by a normalized false-color reference. Projection, norm control, and validation-calibrated injection regulate how much spectral detail is added. Across 86 held-out acquisitions, category balanced accuracy rose from 0.812 to 0.883.

The paper also reports that using the model's own token composition rather than a pooled replacement improves reuse of an external category-derived probe on frozen decoder logits (0.701 vs 0.590). Grouped tests recovered within-category spectral shape with R² values of 0.802 and 0.337 from continuous and posterior-code states, and category-matched spectral swaps roughly doubled the target-specific recovery increment.

The authors conclude that spectral enrichment and preservation of the receiving token work together at the pretrained interface, suggesting a path to extending frozen VLA models to new sensing modalities without architectural changes.