Keep both short property phrases and full tactile descriptions if a touch-language model must answer different kinds of queries. A phrase is easier to compare with an attribute such as surface texture; a sentence can preserve context and change during contact. Touch100k’s paper explicitly distinguishes phrase-level tactile features from sentence-level descriptions that carry richer contextual and dynamic relationships. The practical decision is to choose which level supervises each task, then evaluate the levels separately rather than treating every caption as interchangeable text.
Suppose an editor is labeling GelSight clips from a gripper touching a bottle. “Ribbed surface” would be a short attribute label. “The contact patch moves across the ridges as the bottle turns” would be a contextual description. These are illustrative labels, not quoted Touch100k annotations. If the downstream test retrieves objects by material or texture, the phrase may be enough. If it asks what changed during contact, stripping the sentence down to a noun phrase discards the relationship being tested.
Let the downstream question determine the label level
Touch100k pairs tactile, language and visual information. Its authors introduce TLV-Link, a representation-learning method, and evaluate material-property identification and grasping prediction. Those tests support a representation claim within the reported dataset and setups; they do not demonstrate that a robot controller will improve merely because a caption is longer. Label design should begin with a target question and a held-out test, not with a target word count.
| Reader’s downstream question | Label to preserve | Failure to look for |
|---|---|---|
| “What surface property does this touch suggest?” | A concise phrase for the property, with annotation vocabulary controlled | The model retrieves a material name from the paired image instead of tactile evidence. |
| “What changed while the object moved?” | A sentence that names the contact event and its temporal relationship | A static phrase erases the motion or cause-and-effect cue. |
| “Which caption matches this tactile clip?” | Both levels, tagged separately | A score rises because the same object or scene appears in training and test. |
For a new dataset, store the phrase and sentence as distinct fields tied to the same tactile segment, camera view and time window. Preserve the annotator instructions and examples for each field. A long caption can smuggle in the visual identity of an object that the tactile signal cannot support; a short label can flatten a meaningful slip or deformation sequence. Audit a small blinded sample by asking whether a reviewer could justify each label from the permitted modality. If vision is allowed in the intended task, state that explicitly and test a touch-only baseline to understand what vision contributes.
Split by object, recording session or contact sequence where those identities matter. Adjacent frames from the same touch should not be assigned to both train and test merely to increase the number of pairs. Evaluate phrase retrieval, sentence retrieval and the actual downstream property or grasp-prediction task separately. A higher cross-modal retrieval score is evidence about representation under that split, not a measured increase in grasp success on a physical robot.
Check the artifact and use restrictions before building on it
The Touch100k project page links the dataset resource and the official TLV-Link implementation. The repository labels its code MIT, its dataset CC BY-NC 4.0, and models trained on that dataset as research-only. These are different permissions. A team evaluating a commercial use case should resolve dataset and model rights with the rights holder before reusing them; the code license alone does not settle that question. The linked data files were not inspected or downloaded for this article.
If the recording hardware is not yet fixed, the GelSight Mini and DIGIT 360 comparison addresses that earlier sensor choice. Touch100k asks a later question: whether the language paired with touch retains the semantic detail the downstream task needs. The useful output is a label schema, a modality-aware split and separate metrics for phrase and sentence supervision. No dataset training or robot grasp test was performed for this article.