Amap Releases ABot-M0.5: Unified World-Action Robotic Model

青稞AI · wechat · 2026-07-07

Amap released a new embodied AI model, ABot-M0.5, positioned as a Unified World-Action Model (WAM) to solve the disconnect between prediction and control in mobile manipulation. Built on the Wan2.2 video diffusion backbone, it introduces frame-level "intermediate latent actions" as a spatiotemporal bridge. A dual-layer hybrid Transformer (D-MoT) decouples chassis movement from robotic arm operations, and a "Dream Forcing" training paradigm uses model self-predicted videos to train the action predictor, suppressing error accumulation. It comprehensively outperforms π0 and GR00T-N1.5 on benchmarks like RoboCasa365 and RoboTwin2.0, setting new SOTA in long-horizon and fine-grained control.

Original post →

More from Embodied

Embodied channel →