A robot can locate an object without knowing whether it has a secure grip. Touch sensing addresses that second problem: what happens where the machine actually meets the material. Its promise is not another view of the room, but a closer view of contact.
A camera inside the fingertip
Some tactile sensors turn touch into images. A soft sensing surface changes shape against an object; an internal imaging system reads that deformation. GelSight describes this approach as capturing the surface topography through a conforming elastomer. The camera is looking at the contact surface, rather than watching the object from across a workspace.1
That distinction matters. External vision helps a robot approach the right place. Tactile information can help describe the interaction after contact. Combining the two is more useful than treating touch as a replacement for sight.
The 2020 DIGIT paper demonstrated a compact optical tactile sensor mounted on a multi-fingered robotic hand. The researchers trained model-based controllers to manipulate glass marbles within the hand.2 This was a concrete example of touch feeding movement, not simply producing an attractive sensor image.
Learning what contact means
Reading a tactile image is a separate challenge from capturing it. The same pattern of deformation must become information a robot can use: force, slip, or a change in an object's position.
Sparsh, published in 2024, explored reusable touch representations trained on more than 460,000 tactile images. Its six-task TacBench evaluation compared these representations with models trained end to end for particular tasks. The authors reported improved performance across the benchmark, especially when labeled data was limited.3
The important idea is reuse. A model can learn patterns from many unlabeled contact images before being adapted to a particular prediction task. That could reduce the amount of specialized labeling needed for each application. It does not remove the need to validate the sensor, material and task together.
The underlying experiments also define the result's reach. For force and slip data, the released methods describe three indenter shapes and controlled loading. One pose-estimation dataset used a stationary hand while a person moved the object.4 Predicting those signals is valuable; it is not the same as independently completing an unfamiliar assembly job.
Better fingers are only part of the hand
DIGIT 360 takes a broader hardware approach. Its project documentation describes a compliant, rounded sensing surface, multiple sensing modalities and on-device processing. It also provides drivers and mechanical design files for integration.5
This suggests three distinct layers to evaluate: the sensor captures contact; the model interprets it; the controller decides how to move. A stronger result in one layer can help the others, but does not establish the performance of the complete system.
For a practical trial, start with a contact problem rather than a sensor specification. Does a part slip during rotation? Does handling damage a soft surface? Does alignment become uncertain after the gripper closes? Compare the same task with and without the added feedback. Track successful handling, damage and cycle time, then repeat across the materials and surface conditions the system will encounter. This is BIGFOOT's proposed evaluation approach, not a result reported by the cited experiments.
Research value and maintenance are different questions
There is a useful implementation detail behind the research: the official Sparsh repository was archived on April 1, 2026 and is read-only as of this review.4 Its code and findings remain useful starting points, but availability should not be confused with ongoing upstream maintenance. An integration plan needs explicit responsibility for compatibility and future fixes.
The strongest case for robot touch is therefore specific: it can add information about contact that a distant camera cannot directly measure. The next step is to connect that information to a better action. Start with the handling failure that matters, choose the signal that could reveal it, and test whether acting on that signal improves the outcome.
References
- GelSight. Tactile Sensing Technology. Manufacturer explanation; publication date not stated. Accessed October 17, 2026.
- Lambeta and colleagues. DIGIT: A Novel Design for a Low-Cost Compact High-Resolution Tactile Sensor with Application to In-Hand Manipulation. IEEE Robotics and Automation Letters, 2020; arXiv version submitted May 29, 2020.
- Higuera and colleagues. Sparsh: Self-supervised touch representations for vision-based tactile sensing. CoRL 2024; arXiv version submitted October 31, 2024.
- Meta FAIR and collaborators. Official Sparsh repository and dataset methods. Archived April 1, 2026; repository status and methods checked October 17, 2026.
- Meta FAIR. DIGIT 360 project documentation. Project introduced in 2024; current documentation accessed October 17, 2026. Hardware descriptions are developer claims, not independent field measurements.