Part 7 of our series on the data problem in Vision-Language-Action research. [Part 1: the bottleneck overview · Part 2: the language gap · Part 3: the failure data · Part 4: the evaluation crisis · Part 5: the synthetic trust layer · Part 6: labeling human video]
Here is a statistic that says more than its size suggests: when we split the 1,228 VLA papers into ten chunks for parallel analysis, one chunk of 123 papers contained zero tactile-centric work. Not a little. None. Across the full corpus, audio is thinner still: roughly three efforts in 1,228 papers touch it meaningfully.
For a field built on manipulation, on contact, grip, force, the sound of something seating into place, that absence is strange. It is stranger once you widen the lens to the whole field. In the broader scan of 53,800 robotics papers this corpus was drawn from, manipulation and grasping is the second-most-researched area in all of robotics, with 2,961 papers in the last twelve months alone. Manipulation is, definitionally, contact. And yet the sensory channel that contact actually produces, touch, is almost entirely missing from the data. The field studies grasping intensively and records almost none of what a grasp feels like.
Vision-Language-Action is, in practice, Vision-Language-Action-and-almost-nothing-else. The corpus is overwhelmingly RGB plus language plus proprioception. The modalities most specific to physical manipulation are the ones the field has barely begun to collect. This post is about why that gap exists, why it's larger than it looks, and why it's an opportunity rather than a curiosity.
Contact-rich manipulation without touch is a strange proposition. A parallel-jaw gripper closing on a wine glass, a cable seated into a port, a lid twisted until it catches: these are force-and-contact events that vision resolves poorly and proprioception misses entirely. A camera cannot see the moment a grip begins to slip. Joint encoders cannot tell you whether the object in the gripper is a rigid box or a crushable cup. That information lives in the tactile and force channels, and if you don't record those channels, no model can learn from them, however much video you collect.
The field knows this. It just doesn't have the data. The papers that do tackle touch name the same root cause, and it's a data cause. The scarcity is of aligned vision-tactile-language data: synchronized streams that pair what the robot sees, what it feels, and what it was told to do (arXiv:2605.27886). The blocker to progress on force-aware manipulation is the absence of large multimodal datasets in the first place (arXiv:2507.17294). This is not an architecture problem waiting on a better model. It is the Part 1 thesis in its purest form: the binding constraint is data, and here the data doesn't exist at all.
It connects directly to Part 3, too. The manipulation literature's own diagnosis of its bottleneck is that real grasp attempts are slow and expensive to collect and that most failures go unlabeled. Touch is precisely the signal that would tell you why a grasp failed, whether the fingers never made contact, made contact and slipped, or crushed the object. The missing modality and the missing failure data are the same blind spot seen through two sensors.
The tell that a modality is greenfield is that merely collecting it is a headline result. In mature areas, a new dataset is incremental; in tactile and audio robotics, a dataset is the contribution.
Look at the papers that exist. "HapTile: A Haptic-Informed Vision-Tactile-Language-Action Dataset for Contact-Rich Imitation Learning" announces its contribution in its title, and the contribution is the dataset: it exists to supply the aligned vision-tactile-language data the field lacks. "Multi-Resolution Tactile Imitation Learning for Contact-Rich Robotic Manipulation" is doing the same foundational work one modality over. Alongside them, force-aware policies like ForceVLA, vision-tactile-language models like OmniVTLA, and audio-inclusive systems like RoboOmni are all flagged as flagship work for one reason: the data doesn't otherwise exist. Each is publishable primarily because it built the aligned multimodal dataset that was missing, and each is small relative to the RGB corpora that dwarf them. When gathering the data is itself the novel act, you are looking at a field at the beginning of its data curve, not the middle.
Audio deserves its own note, because it is the most neglected channel of all. Three efforts across 1,228 papers is not a research area; it is a rounding error. Yet sound is a rich, cheap signal for exactly the contact events vision struggles with: the click of a connector seating, the scrape of a mis-aligned part, the rattle that says an object is loose in the gripper. A contact microphone costs almost nothing and streams continuously. The barrier is not the sensor. It is that no one has built the aligned, labeled audio-vision-action datasets that would let a model use the signal, which is the same barrier as touch.
Multimodal manipulation data is hard to produce for reasons that compound. It requires specialized sensors (tactile skins, force-torque sensors, contact microphones) that most collection rigs simply don't have. The streams must be captured in physical synchrony, because a tactile reading is meaningless unless you know which visual frame and which action it corresponds to. And it can't be scraped or scaled from the web the way image-text pairs can; there is no internet of aligned touch data. Synthetic generation, as Part 5 argued, struggles here too: simulating faithful contact physics and tactile signal is far harder than rendering plausible RGB, so the synthetic shortcut that is starting to work for vision is not yet available for touch.
Every one of those difficulties is also a moat. The RGB-plus-language space is crowded and increasingly commoditized, with hundreds of new VLA papers a year competing on the same visual data. The multimodal frontier is wide open precisely because the data is hard to build, which means the teams that build the aligned vision-tactile-language and audio corpora now will define the datasets everyone else benchmarks against later. First-mover advantage in a data-scarce modality is durable in a way it never is once the web-scraped floodgates open, because here there are no floodgates to open. The data can only be produced deliberately, by instrumented collection and human annotation, and whoever produces it first owns the reference set.
And multimodal annotation is a natural extension of the same alignment work this whole series has been about. Segmenting and labeling a tactile or audio stream, aligning it to vision, language, and action, grounding "you'll feel it click when it's seated" to the moment the force signal spikes: this is the intention-aligned, fine-grained annotation from Parts 2 and 6, extended to the senses the field has ignored. The failure taxonomy from Part 3 extends here too, because a force trace is often the clearest record of the exact instant a manipulation went wrong.
You don't need to act on tactile and audio data today to take the signal seriously. The point is directional. The field's next expansion is not more RGB manipulation data, of which there is plenty and of which the marginal value is falling (see Part 8). It's the modalities that contact-rich manipulation actually depends on and currently lacks, in a field where manipulation is the second-largest research area and touch is nearly absent from the record.
For teams thinking a year ahead, the move is concrete. Instrument for touch, force, and sound now, while almost no one else is. Capture the streams in hardware synchrony with your existing vision and action logs, because unaligned tactile data is nearly as inert as unlabeled video. Treat the annotation schema for those streams, what to segment, what to label, how to tie a felt event to a language description, as the durable asset rather than the sensor readings themselves. And borrow the lesson from every prior part: automate what you can, but put human judgment on the fine-grained, causal labels that generic models can't produce.
A chunk of 123 papers with zero tactile work is not a description of a dead end. It's a map of open ground, in the part of the field that matters most.
Next in the series, the capstone: what the scaling debate actually shows, and why quality beats quantity.
This post draws on our analysis of 1,228 VLA papers from arXiv cs.RO (Feb 2023 – Jun 2026), part of a broader scan of 53,800 robotics papers. See Part 1 for methodology.