ABOT-M0.5 Unified Mobility-and-Manipulation World Action Model

AMAP CV Lab
July 2026
Code Paper BibTeX

Overview

Mobile manipulation remains challenging for embodied learning: VLA policies are typically reactive and lack explicit world modeling, while existing World Action Models (WAMs) are poorly aligned with mobile manipulation in temporal granularity, action space, and train-test consistency. ABot-M0.5 is a unified WAM that closes these gaps with intermediate latent actions, a dual-level Mixture-of-Transformers architecture, and Dream Forcing training.

ABot-M0.5 teaser

Temporal Alignment

Intermediate latent actions bridge coarse video prediction and fine-grained robot control via local visual state transitions.

Action Disentangling

Dual-level MoT separates modality streams and mobility vs. manipulation subspaces to reduce action conflicts.

Dream Forcing

Inverse dynamics is trained on model-predicted videos, improving train-test consistency for long-horizon rollouts.

46.6%
RoboCasa365
94.1%
RoboTwin 2.0
99.4%
LIBERO
83.4%
LIBERO-Plus

Method

ABot-M0.5 model architecture

Model architecture of ABot-M0.5.

Intermediate Latent Actions

Instead of mapping video chunks directly to robot controls, ABot-M0.5 introduces frame-level intermediate latent actions that capture local visual transitions. This forms a three-stage cascade: world modeling → motion abstraction → control decoding. Latent actions are embodiment-agnostic and can be learned from large-scale video data, then decoded into embodiment-specific controls.

Intermediate latent actions

Intermediate latent actions bridge video latents and executable controls.

Dual-Level Mixture-of-Transformers

Mobile manipulation couples low-frequency global base motion with high-frequency arm contact dynamics. ABot-M0.5 uses a Dual-Level Mixture-of-Transformers (D-MoT) to disentangle both modality representations (video / latent action / executable action) and heterogeneous action subspaces (mobility vs. manipulation), while joint attention preserves cross-stream coordination.

Dual-level Mixture-of-Transformers

Dual-level MoT disentangles modalities and action subspaces.

Dream Forcing

Standard teacher forcing trains inverse dynamics on ground-truth future videos, but autoregressive inference must act on self-predicted videos. Dream Forcing closes this gap with a two-phase forward pass: first generate dreamed video latents, then train action prediction on those predicted videos—making training conditions match deployment.

Dream Forcing 1
Dream Forcing 2

Dream Forcing aligns train-time inverse dynamics with autoregressive inference.

3-Stage Separate & Progressive Training

A structured attention mask with disjoint temporal/modal indicators prevents modality leakage across video, latent-action, and action streams. Training proceeds progressively: large-scale video pretraining → supervised fine-tuning with teacher forcing → Dream Forcing alignment for robust long-horizon rollouts.

Attention mask

Structured attention masking for 3-stage separation.

Attention maps

Attention maps become more target-focused after pretraining and fine-tuning.

Results

ABot-M0.5 achieves strong performance on mobile manipulation, fine-grained manipulation, and real-robot deployment.

Benchmark overview

Benchmark overview of ABot-M0.5.

RoboCasa365 (Mobile Manipulation)

Under pretraining evaluation, ABot-M0.5 reaches 46.6% average success (Atomic-Seen 75.9%, Composite-Seen 38.3%), outperforming GR00T-N1.5 (23.9%), RLDX-1 (36.0%), and Qwen-RobotManip (35.9%). In the Target 100% setting, it achieves 54.2% average success.

RoboCasa365 results

Results on RoboCasa365 mobile manipulation.

RoboTwin 2.0

ABot-M0.5 reaches 94.0% on Clean (Easy) and 94.2% on Randomized (Hard), for an average of 94.10%.

RoboTwin 2.0 results

Results on RoboTwin 2.0.

LIBERO & LIBERO-Plus

On LIBERO, ABot-M0.5 achieves 99.4% average success across Spatial / Object / Goal / Long. On LIBERO-Plus zero-shot robustness, it reaches 83.4% total average under camera, robot, light, and noise perturbations.

LIBERO results

Results on LIBERO / LIBERO-Plus.

Real-Robot Deployment

Deployed on an Agilex Piper (6-DoF) mobile manipulation platform with only 50 real demonstrations per task. ABot-M0.5 achieves strong fine control (Peg Cylinder: 70% success / 96% process score) and long-horizon multi-stage tasks (Organize Plate 70%, Arrange Fruits 80%, Cup Stacking 80%, Arrange Flower 60%).

Real robot results

Real-world mobile manipulation deployment.

Real-Robot Demos

Three real-world mobile manipulation demos on an Agilex Piper platform.

Cup Stacking demo

Cup Stacking

Arrange Fruits demo

Arrange Fruits

Arrange Flower demo

Arrange Flower

Author Team

Author contributions in the following areas are as follows:

  • Data Collection & Standardization: Yandan Yang, Ronghan Chen, Yuzhi Chen, Haoyun Liu, Dekang Qi
  • Model & Training: Ronghan Chen, Zuojin Tang, Tong Lin, Yandan Yang
  • Post-Training & Evaluation: Zuojin Tang, Tianlun Li, Haoning Wu, Ronghan Chen, Tong Lin, Mingxin Wang, Bin Hu
  • Real-Robot Experiments & Deployment: Dongjie Huo, Lulu Zheng, Botai Yuan
  • Writing: Yandan Yang, Ronghan Chen, Zuojin Tang, Dekang Qi, Haoyun Liu
  • Challenge Submission: Yanqing Zhu‣, Wei Mei, Yuze Xuan, Haolong Yang, Dongjie Huo
  • Project Lead: Xinyuan Chang
  • Advisor: Mu Xu†, Zhiheng Ma

† Corresponding Author: xumu.xm@alibaba-inc.com  ·  ‣ Lead for Challenge Submission  ·  AMAP CV Lab

Citation

If you find our work helpful, please cite: 

@article{chen2026abotm05,
  title={ABot-M0.5: Unified Mobility-and-Manipulation World Action Model},
  author={Chen, Ronghan and Yang, Yandan and Tang, Zuojin and Huo, Dongjie and Lin, Tong and Wu, Haoning and Liu, Haoyun and Chen, Yuzhi and Zheng, Lulu and Yuan, Botai and Li, Tianlun and Wang, Mingxin and Qi, Dekang and Hu, Bin and Mei, Wei and Xuan, Yuze and Yang, Haolong and Zhu, Yanqing and Xu, Mu and Ma, Zhiheng and Chang, Xinyuan},
  journal={arXiv preprint arXiv:2607.00678},
  year={2026}
}