AnchorWorld

AnchorWorld is a research framework from an internship at the Kling Team, Kuaishou Technology, for egocentric (first-person) interactive world simulation. It uses 3D human motion as the action signal, so an agent or player can navigate and interact with a generated world, and it renders the result as first-person video, including object interactions.

To handle body parts that are out of view in first-person footage, it adds third-person video as auxiliary training supervision, using a projection-based action representation that combines full-body motion with camera trajectories. The model is pre-trained on third-person videos and then adapted to the egocentric viewpoint.

World customization works through anchor views. Each anchor view has an RGB image for appearance, a 3D pose for spatial grounding, and an evolution prompt describing how the local scene changes over time. These are injected into a pre-trained video model without architectural changes: in-context image tokens, pose embeddings, and masked cross-attention for the prompts.

The project page compares it against PlayerOne (re-implemented on Wan2.2 TI2V 5B) and CaM. At the time of the source material, code was reported to be under review, and no weights or public demo were listed. Hand poses are excluded because their estimation from egocentric data is unreliable.