Researchers from Meta’s Fundamental AI Research (FAIR) team have introduced new tools and collaborations to enhance tactile sensing, robotic dexterity, and human-robot interaction. These advancements bring robots closer to functioning as practical partners, bridging physical and virtual environments with greater accuracy.

Advancing Touch Perception: Meta Sparsh and Digit 360

One core aspect of human interaction is touch—a skill most robots lack. Meta has responded with Meta Sparsh, a versatile touch-sensing model built to work across various sensors and tasks. Developed through self-supervised learning, Sparsh relies on over 460,000 tactile images to interpret touch without manual labels. This breakthrough paves the way for precision in sectors where touch sensitivity is critical, such as healthcare and manufacturing.

Complementing Sparsh is Digit 360, an advanced tactile sensor with capabilities that include detecting slight spatial shifts, vibrations, and even surface temperature. With over 8 million “taxels,” Digit 360 captures detailed touch interactions, mirroring the sensitivity of human fingertips. This sensor is expected to reshape touch-based AI research by enabling robots to understand and respond to the physical world with more accuracy.

Meta Digit Plexus: A Platform for Robotic Dexterity

To enable intricate robot manipulation, Meta has introduced the Digit Plexus platform, a comprehensive solution for integrating tactile sensors like Digit 360 into robotic hands. This platform standardizes tactile data collection and analysis, offering researchers a streamlined way to develop advanced robotic dexterity. With potential applications across medical devices, VR, and manufacturing, Digit Plexus enables robots to handle objects with near-human dexterity.

Meta’s collaboration with GelSight Inc. is expanding access to Digit 360, with mass production slated for next year. Additionally, a partnership with Wonik Robotics will bring about a new version of the Allegro Hand, fully integrated with Digit Plexus. These partnerships aim to accelerate tactile AI research and enhance practical applications, making advanced robotic manipulation a more accessible reality.

PARTNR Benchmark: Improving Human-Robot Teamwork

FAIR’s PARTNR Benchmark tackles the social and collaborative dimensions of robotics. Built on the Habitat 3.0 simulator, this benchmark enables large-scale evaluations of AI planning and reasoning within human-robot collaboration scenarios, preparing robots for realistic interactive tasks. In a simulated home-like environment, PARTNR provides a scalable, safe way for robots to practice collaborative tasks.

With a database of over 100,000 natural language tasks, 60 varied home settings, and thousands of distinct objects, PARTNR allows researchers to rigorously test AI models for human-robot interaction. Early testing has highlighted specific challenges AI planners face in coordinating and recovering from task errors, offering valuable insights to guide future improvements in collaborative AI.

Moving Toward Intelligent Robotic Partners

These advancements in AI-driven touch perception, dexterity, and collaboration are significant steps toward intelligent, embodied robots capable of meaningful interactions. By prioritizing open-source tools and standardized benchmarks, Meta’s FAIR team is encouraging the AI community to build upon this work and contribute to the development of practical, human-centered robotics.

From advanced prosthetics and assistive robots to immersive VR experiences, these technologies are gradually integrating AI and robotics into everyday life.

Through collaborative innovation, the future of human-robot partnerships is increasingly within reach.