ABot-M0.5: Unified Mobility-and-Manipulation World Action Model

Mobile manipulation requires robots to navigate cluttered spaces while performing precise object interactions. Existing Vision-Language-Action (VLA) models often lack explicit world modeling, while current World Action Models (WAMs) suffer from structural mismatches: they use coarse video chunks that obscure fine-grained contact dynamics, entangle navigation and manipulation actions, and exhibit train-test distribution shifts during autoregressive rollout. This paper asks: how can we align world modeling, action abstraction, and deployment-time behavior to enable robust, long-horizon mobile manipulation?

ABot-M0.5 addresses these bottlenecks through three core design principles:

• Temporal Alignment: It introduces intermediate latent actions that bridge the gap between coarse video latents and fine-grained robot controls, capturing local visual transitions that are otherwise lost.
• Action-Space Alignment: It employs a Dual-Level Mixture-of-Transformers (D-MoT) architecture that disentangles modality-specific representations and separates heterogeneous action subspaces (base mobility vs. arm manipulation), preventing gradient interference.
• Train-Test Alignment: It introduces a Dream-Forcing training strategy, where the action predictor is trained on self-dreamed future video latents rather than ground-truth futures. This exposes the model to the same imperfect visual conditions it encounters during inference, significantly reducing exposure bias.

ABot-M0.5 achieves state-of-the-art results across several challenging benchmarks, including RoboCasa365, RoboTwin 2.0, and LIBERO-Plus. It demonstrates superior long-horizon task success and fine-grained control accuracy compared to leading VLA and WAM baselines. Real-world experiments on an Agilex Piper arm show that the model maintains high performance in both precise insertion tasks and complex multi-stage manipulation, validating its transferability and robustness.