Robots are learning contact-rich manipulation — insertion, deburring, fixture loading — and the hardest signal to find is measured force. We record it where skilled work actually happens: across five-plus industrial sites in India — machine-tool plants, warehouses, shipping yards — under exclusive collection partnerships.
Episodes of skilled machinists doing paid production work — manual deburring of ground castings, part loading into fixtures, go/no-go gauge inspection — captured with handheld instrumented grippers based on Stanford's open-source UMI-FT platform.
Independent published results report large gains on contact-rich tasks when measured force joins a policy's inputs — while most of today's manipulation data carries no force signal at all, and leading force-aware models train on wrench estimated from joint torques. We collect the measured signal, calibrated at the fingertip, from work where force is the skill.
Rigs are in fabrication. Scoped collection programs — your task families, modalities, and volumes — are open for contract now. Our first public sample dataset — force-annotated contact tasks, with a published task suite and scoring protocol — is planned for October 2026.
Vendor and pilot inquiries welcome now.
ryan@preloaddata.com