Alibaba Uses World Models for Digital Teleoperation

量子位 · wechat · 2026-07-17

Alibaba DAMO Academy proposed RynnWorld-Teleop, shifting from "human teleoperating a real robot" to "human gestures driving a generative world model." The system directly synthesizes first-person robot video in a digital world and automatically generates joint-level action labels.

The article details three key conditions and implementations of this digital teleoperation: first-person robot perspective, traceable actions, and real-time interaction. The corresponding tech includes depth-aware skeleton rendering, progressive cross-domain training, and distilling a bidirectional video diffusion model into a causal student model running at 40FPS in real time.

In experiments, data generated by RynnWorld-Teleop can directly train real robot policies. For bimanual dexterous hand tasks, usable policies were achieved with zero real-world data. When synthetic data was mixed with real-world data, success rates improved further across most tasks, proving the synthetic trajectories closely match real data in both pixels and feature distribution.

Original post →

More from Embodied

Embodied channel →