Robots Are Learning to Feel
New tactile datasets could solve a longstanding challenge in robotics
By Edd Gent, Contributing Editor, IEEE Spectrum
For years, robots have excelled at seeing and navigating the world, but they have lagged behind in one crucial sense: touch. While computer vision has advanced rapidly, tactile sensing has remained one of robotics' most stubborn challenges. Now, new datasets โ including T-Rex โ are helping robot hands perform fiddly manipulation tasks with greater success, signaling a turning point for the field.
Why Touch Is So Hard for Robots
Vision and language have dominated AI progress because they benefit from massive, easily collected datasets. Touch, by contrast, is intimate, high-dimensional, and difficult to instrument. Every contact event involves force, texture, temperature, and slip โ all changing in milliseconds. Unlike a photograph, a tactile signal is inseparable from the specific hardware that produced it, making datasets hard to share and generalize.
The T-Rex Dataset and the New Wave of Tactile Data
T-Rex (Tactile-Reactive Dexterous manipulation) is one of a growing number of datasets designed to close this gap. By recording rich tactile signals alongside motor actions, T-Rex gives robot hands the sensory grounding they need to handle objects that slip, deform, or require delicate force control โ tasks like plugging in a cable, sorting produce, or assembling small components.
What makes these datasets promising is scale and standardization. As more labs contribute tactile data in compatible formats, robot learning models can begin to generalize across hardware, much like vision models did after ImageNet.
2026 Context: Tactile Sensing Moves Toward Maturity
By 2026, tactile sensing is transitioning from research curiosity to commercial necessity. Humanoid robot programs, warehouse automation, and advanced manufacturing are all demanding hands that can feel what they're doing. Falling costs of high-resolution tactile skins, combined with foundation models that can learn from multimodal data, are accelerating this shift. The question is no longer whether robots can feel โ it's how quickly they can learn to feel well enough to work alongside us.
The Road Ahead
Challenges remain: data collection is labor-intensive, tactile hardware is still fragmented, and sim-to-real transfer for touch is far less mature than for vision. But the direction is clear. As datasets like T-Rex mature and proliferate, robots may finally gain the sense that makes true dexterity possible โ the sense of touch.
T-Rex project page: tactile-reactive-dexterous.github.io