Event

FTP-1 General-Purpose Tactile Foundation Policy Released

1 reports · 1 sources

The research team released FTP-1, a general-purpose foundation tactile policy pretrained across tactile sensors and embodiments. It uses heterogeneous encoders to map image, array, and state-type tactile signals into unified morphology-aware latent tokens, which are then jointly modeled by shared tactile Transformer experts. According to the authors, FTP-1 was pretrained on approximately 3,000 hours of tactile manipulation data, aggregated from 26 data sources and covering human and robot demonstrations on 21 sensor types. In downstream fine-tuning experiments across 5 hardware configurations, it improved contact-intensive manipulation by 17.2% on seen sensor configurations; on two previously unseen tactile sensor configurations, it improved the success rate by 31%. The pretrained model, dataset, and training code have been made publicly available.