Meta FAIR Lab recently introduced a groundbreaking technology called "Sparsh," which is a human-like artificial multimodal fingertip haptic perception system. This technology has the potential to revolutionize the field of robotic manipulation by granting robots near-human tactile sensing capabilities.
The Sparsh model employs self-supervised learning, utilizing over 460,000 tactile images for pre-training without the need for manual data annotation, enabling it to learn universal tactile representations.
The model supports various types of visual-tactile sensors, including DIGIT, GelSight2017, and GelSight Mini, significantly enhancing the performance of robots in tactile sensing tasks such as force estimation, slip detection, pose estimation, grasp stability prediction, and fabric recognition. Researchers have also constructed a standardized benchmark testing platform called TacBench to evaluate the performance of different tactile sensors and models across various tasks.
Test results show that the Sparsh model excels in all six tasks on TacBench, especially when data is limited, outperforming traditional task-specific and sensor-specific models. For instance, in force estimation and slip detection tasks, Sparsh can achieve satisfactory results even with only 1% of the labeled data. This indicates that Sparsh can help robots better understand the physical properties of objects and perform more precise manipulations.
The release of the Sparsh model marks a significant breakthrough in the field of AI tactile perception. In the future, with the accumulation of more data and further optimization of the model, Sparsh is expected to transform the way robots interact with the physical world, driving the application of robotic technology in broader areas.
Paper Link:
https://scontent-sjc3-1.xx.fbcdn.net/v/t39.2365-6/464969941_1107633400780143_7479102347328147009_n.pdf?_nc_cat=103&ccb=1-7&_nc_sid=3c67a6&_nc_ohc=y8Ui1HEw3BQQ7kNvgFe-ePu&_nc_zt=14&_nc_ht=scontent-sjc3-1.xx&_nc_gid=AeaFsuZziasVwPfMQsEoZqu&oh=00_AYAMqxGq0ATCySDxZWB0ZT8BgSkogYmj13c9f3ytVtkmSg&oe=672DEEE4