TAPESTRY: A Single Backbone for Video Generation and Robot Control
TAPESTRY: A Single Backbone for Video Generation and Robot Control
TAPESTRY is a world action model that repurposes a pretrained video diffusion transformer into a robotic policy by treating robot actions as single-patch tokens in the video model's latent space. The model generates compact action token sequences in a single DDIM step, achieving smooth action chunks at 84 ms per action chunk on consumer hardware, with strong performance across simulation and real-robot experiments.
TAPESTRY reuses video priors for action-only inference and reports latency plus simulation and physical-robot evaluation. A specific world-action-model design to test; the abstract is not independent replication.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org