Contribution of IMPACT’s Language-Conditioning Paths

Determine the individual contributions of IMPACT’s projected text-token pathway and FiLM-based visual-modulation pathway to language-based task selection and manipulation performance.

Background

IMPACT conditions its policy through two mechanisms: projected instruction tokens appended to the transformer sequence and instruction-derived FiLM coefficients applied to visual features. The instruction-shuffling experiment evaluates the combined policy but does not isolate either mechanism, leaving their individual effects unresolved.

References

This intervention supports language-based selection among familiar goals; generalization to unseen wording and the individual contributions of the two conditioning paths remain untested.

— vla.simd: Efficient CPU Inference for Language-Conditioned Manipulation  (2609.24274 - Nguyen et al., 21 Sep 2026) in Section 5, paragraph "Instruction following"