VLA에게 부족한 결정적 감각 | 로봇에게 Force·Tactile이 반드시 필요한 이유 (Why VLAs Need Force and Tactile Sensing for Robot Manipulation)

TL;DR

A Korean-language explainer video from the “엥지유니버스 | 로봇 엔지니어” channel surveying recent attempts to add force and tactile sensing to VLA (Vision-Language-Action) models. The argument: VLA models are strong on the vision + language + action stack, but vision alone cannot tell a robot whether an object is slipping, whether grip force is too high, or whether contact is misaligned — and contact-rich manipulation (plugging in cables, holding fragile objects, deformable-object handling) depends on those signals. The video groups the literature into three architectural patterns for fusing force/tactile with VLA, plus a second framing — touch-as-a-language — where tactile signals convey scene properties (softness, hardness, slip, pressure) that aren’t visible from RGB. No transcript was retrievable at filing time, so this page captures the structure rather than per-paper claims; per-paper details should be added if/when a transcript or accompanying writeup surfaces.

Key claims

Pure VLA models (vision + language + action) leave a gap on contact-rich manipulation because vision can’t report slip, grip-force magnitude, resistance, or contact misalignment [video description / sharer TL;DW].
Recent VLA + force/tactile work falls into three architectural patterns: (a) inject force/tactile directly into the VLA model, (b) deeply fuse tactile with vision and language inside the policy, and (c) keep the VLA mostly unchanged and wrap it with tactile feedback around the planning/control loop [video framing per sharer TL;DW].
A second direction in the surveyed work treats touch as its own “language” the robot reads, conveying scene properties (softness, hardness, slipping, pressure, misalignment) that are hard to recover from images alone [video framing per sharer TL;DW].
Net argument: VLAs unlocked vision and language for robots, but for real-world manipulation they probably also need first-class physical senses (force, touch) — i.e. the sensorimotor axis is not optional [video conclusion per sharer TL;DW].

Method

Not a research artifact — a survey/explainer video on a Korean robotics-engineering channel. The published title translates roughly as “The decisive sense VLA lacks | Why robots absolutely need Force·Tactile.” Filed because it organizes the multi-paper VLA + force/tactile landscape into a small set of architectural patterns and a “touch-as-language” framing, both of which are useful crosscuts for the wiki’s existing VLA cluster. No transcript was available at fetch time, so the claims above are sourced from the video description and the sharer’s TL;DW summary in the originating Slack post.

Results

Not applicable — no quantitative results. The video does not cite specific benchmarks or per-paper numbers in its description; the per-paper evidence base lives in the underlying papers it surveys (which include filed entries like Tactile-Reactive Dexterous Hand: High-Frequency Physical Interaction and PressureVision++: Estimating Fingertip Pressure from Diverse RGB Images on the wiki).

Why it’s interesting

This is the first filed survey of the VLA + force/tactile design space on the wiki — every prior filing has been a single paper or project page. The three-pattern taxonomy (force-inside-the-model vs deeply-fused vs tactile-wrapper) sharpens the question that Tactile-Reactive Dexterous Hand: High-Frequency Physical Interaction (variable-rate MoT with first-class tactile stream) and PressureVision++: Estimating Fingertip Pressure from Diverse RGB Images (infer contact from vision rather than ingest touch) currently sit on opposite ends of: where exactly should tactile enter the VLA stack? It also directly answers the open question listed on the VLA Models concept page — “how should tactile / contact dynamics be incorporated into the VLA stack — as an extra modality conditioning the VLM, as a separate low-level controller below the VLA, or via simulation-augmented training?” — by recognizing that the literature has already split into roughly those three camps. The touch-as-a-language framing is a useful complement to Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level‘s sensorimotor-axis counter-position and to the Scaling Robotic Manipulation via Structured World Models and Tactile Sensing — Yunzhu Li (Montreal Robotics) talk (Yunzhu Li, Montreal Robotics) that argued structured world models + tactile are the missing axes in scaling manipulation.