<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-touch">
    <titleInfo>
        <title>Touch What You Say: Contact-Local Tactile Grounding for Physical-Property Instructions</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Le</namePart>
        <namePart type="family">Wang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Kelan</namePart>
        <namePart type="family">Peng</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Zhang</namePart>
        <namePart type="family">Chenyu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Wenjuan</namePart>
        <namePart type="family">Han</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-10</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Hung-Yi</namePart>
            <namePart type="family">Lee</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Sunipa</namePart>
            <namePart type="family">Dev</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jessy</namePart>
            <namePart type="family">Li</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Desmond</namePart>
            <namePart type="family">Elliott</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Budapest, Hungary</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-459-0</identifier>
    </relatedItem>
    <abstract>While vision-language-action (VLA) models excel at robotic manipulation, they frequently fall short when interpreting physical-property language, where adjectives such as fragile or coarse-grained specify constraints that are often visually ambiguous and should modulate behavior primarily during active contact. We therefore propose Contact-Local Semantic Actuation (CLSA), an architecture that extracts semantically structured tactile representations through a language-guided bottleneck and injects them through a contact-gated residual path, confining physical adjustments to contact-relevant stages without perturbing stable nominal motion. To probe instruction-conditioned behavior under matched contexts, we introduce the Action-Property-Constraint (APC) hierarchy. On the language side, APC contrasts action-only (L0), property-aware (L1), and constraint-explicit (L2) prompts; on the trajectory side, offline evaluation uses the corresponding demonstrated behavior strata defined by the data protocol. Across matched offline comparisons, a 140-trial comparative real-robot benchmark, and a separate 240-rollout APC sweep on four tactile-sensitive tasks, CLSA improves instruction sensitivity while maintaining action accuracy. We also report visually ambiguous black-bag trials within the comparative benchmark as a diagnostic of touch-based material recognition after contact.</abstract>
    <identifier type="citekey">wang-etal-2026-touch</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.1120/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>25635</start>
            <end>25648</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Touch What You Say: Contact-Local Tactile Grounding for Physical-Property Instructions
%A Wang, Le
%A Peng, Kelan
%A Chenyu, Zhang
%A Han, Wenjuan
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F wang-etal-2026-touch
%X While vision-language-action (VLA) models excel at robotic manipulation, they frequently fall short when interpreting physical-property language, where adjectives such as fragile or coarse-grained specify constraints that are often visually ambiguous and should modulate behavior primarily during active contact. We therefore propose Contact-Local Semantic Actuation (CLSA), an architecture that extracts semantically structured tactile representations through a language-guided bottleneck and injects them through a contact-gated residual path, confining physical adjustments to contact-relevant stages without perturbing stable nominal motion. To probe instruction-conditioned behavior under matched contexts, we introduce the Action-Property-Constraint (APC) hierarchy. On the language side, APC contrasts action-only (L0), property-aware (L1), and constraint-explicit (L2) prompts; on the trajectory side, offline evaluation uses the corresponding demonstrated behavior strata defined by the data protocol. Across matched offline comparisons, a 140-trial comparative real-robot benchmark, and a separate 240-rollout APC sweep on four tactile-sensitive tasks, CLSA improves instruction sensitivity while maintaining action accuracy. We also report visually ambiguous black-bag trials within the comparative benchmark as a diagnostic of touch-based material recognition after contact.
%U https://aclanthology.org/2026.emnlp-main.1120/
%P 25635-25648

Markdown (Informal)

[Touch What You Say: Contact-Local Tactile Grounding for Physical-Property Instructions](https://aclanthology.org/2026.emnlp-main.1120/) (Wang et al., EMNLP 2026)

• Touch What You Say: Contact-Local Tactile Grounding for Physical-Property Instructions (Wang et al., EMNLP 2026)

ACL

• Le Wang, Kelan Peng, Zhang Chenyu, and Wenjuan Han. 2026. Touch What You Say: Contact-Local Tactile Grounding for Physical-Property Instructions. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 25635–25648, Budapest, Hungary. Association for Computational Linguistics.