Meta FAIR goes all-in on robot touch with Digit 360 and Sparsh

Meta FAIR just made high-resolution robot touch accessible to anyone working in embodied AI. The open release covers Sparsh, a general-purpose vision-based tactile encoder, the Digit 360 multimodal fingertip sensor, the Digit Plexus integration platform, and PARTNR, a benchmark for human-robot interaction.

Robots often struggle with physical manipulation because cameras get occluded the moment fingers touch an object. Meta aims to turn digitized touch into a plug-and-play foundation modality, giving robot hands the ability to feel forces, texture, and slip without custom sensor tuning.

How it works

Vision-based tactile sensors usually vary wildly in gel geometry, internal illumination, and markers, forcing teams to handcraft sensor-specific models with painful data collection. Sparsh bypasses manual force labels by using self-supervised learning.

Meta pre-trained the Sparsh model family on over 460,000 tactile images. On a six-task benchmark spanning physical perception, tactile property comprehension, and dexterous planning, Sparsh outperformed task- and sensor-specific baselines by an average of over 95%.

On the hardware side, Digit 360 provides artificial fingertip sensing with over 18 multimodal features. It features a custom wide-field optical design with more than 8 million taxels and detects contact forces as tiny as 1 millinewton.

To wire everything together, Meta built Digit Plexus, a standardized hardware and software framework that unifies fingertip and tactile skin signals across an entire robotic hand. For system-level reasoning, the open PARTNR benchmark evaluates LLM planners during collaborative human-robot home scenarios.

Availability and hardware

The code, pre-trained weights, and benchmark suites are open-sourced for the research community.

For physical hardware, GelSight is manufacturing and distributing the commercial Digit 360 next year, with researchers able to apply for early access through an open call for proposals. Wonik Robotics is manufacturing the next-generation Allegro Hand with integrated Digit Plexus tactile sensing, also slated for commercial release next year.

Why it matters

If you work on robotic grasping or embodied agents, this is worth testing immediately. Touch data has always been fragmented, and Sparsh provides the first real universal backbone that works across different tactile sensors without labeling headaches.

The catch is physical availability. While the models and simulation benchmarks are ready today, true validation requires high-end tactile hardware that will not ship at scale until next year. Still, Meta's strategy here is clever: by open-sourcing the representation layer and outsourcing manufacturing to GelSight and Wonik, it is laying the groundwork to become the default software stack for dexterous robot manipulation.