ContactWorld shows point-cloud + tactile boosts robot manipulation to 36.1%

ContactWorld, introduced by Zhiyuan Zhang and eight co-authors, is a systematic empirical study and benchmark designed to answer what makes vision-tactile world models effective for contact-rich manipulation. Across 12 tasks — including insertion, disassembly, screwing, and exploratory interaction — the team tested various sensory representations. Their key finding: spatially structured and temporally continuous representations consistently outperform others. Point-cloud observations raised average planning success from 20.7% (wrist-view) or 22.0% (front-view) to 32.1%. Adding tactile force-field representations, which preserve richer spatial structure and interaction dynamics, pushed performance further to 36.1%.

The research also highlights that tactile sensing's effectiveness depends crucially on cross-modal compatibility, not just on scaling the number of sensors. Under long-horizon planning objectives, where prediction errors and contact uncertainty accumulate, tactile input becomes increasingly important. These insights provide concrete guidance for roboticists building world models: prioritize representation structure and multimodal alignment over raw sensor quantity. The work is particularly relevant for industrial automation, assembly, medical robotics, and any domain requiring precise, sustained physical contact with objects.