Robotics / Training / Generation
ABot-M0.5: Unified Mobility-and-Manipulation World Action Model
** Ronghan Chen, Yandan Yang, Zuojin Tang, Dongjie Huo, Tong Lin, Haoning Wu, Haoyun Liu, Yuzhi Chen, Lulu Zheng, Botai Yuan, Tianlun Li, Mingxin Wang, Dekang Qi, Bin Hu, Wei Mei, Yuze Xuan, Haolong Yang, Yanqing Zhu, Mu Xu, Zhiheng Ma, Xinyuan Chang
ABot-M0.5: Unified Mobility-and-Manipulation World Action Model
Authors: Ronghan Chen, Yandan Yang, Zuojin Tang, Dongjie Huo, Tong Lin, Haoning Wu, Haoyun Liu, Yuzhi Chen, Lulu Zheng, Botai Yuan, Tianlun Li, Mingxin Wang, Dekang Qi, Bin Hu, Wei Mei, Yuze Xuan, Haolong Yang, Yanqing Zhu, Mu Xu, Zhiheng Ma, Xinyuan Chang
arXiv ID: 2607.00678
Problem: Existing World Action Models fail at mobile manipulation due to three structural misalignments - coarse video prediction doesn't match fine-grained control, entangled navigation/manipulation action spaces cause gradient interference, and training on ground-truth futures doesn't generalize to the model's own noisy rollouts at inference time.
Key Methodology:
- Intermediate Latent Actions - frame-level, embodiment-agnostic motion representations that bridge coarse video latents and robot actions via a 3-stage cascade (video → latent action → action), enabling fine-grained contact dynamics recovery
- Dual-Level Mixture-of-Transformers (D-MoT) - disentangles modalities (video, latent action, action) and action subspaces (base movement vs. arm manipulation) with separate FFNs/prediction heads, while preserving cross-subspace coordination via joint self-attention
- Dream Forcing - a two-phase training strategy that conditions action prediction on the model's own self-dreamed (predicted) video latents instead of ground-truth futures, eliminating the train-test distribution gap
Key Results:
- RoboCasa365: 94.0% success rate with the full 3-stage pipeline vs. 87.6% for direct video-to-action mapping and 91.1% for channel-concatenated baselines
- Dream Forcing: boosts success rate from 67.55% → 70.56% (+3.01% absolute) in just 5k training steps, while continued teacher forcing degrades to 66.78%
- RoboCasa365 pretraining advantage: 49.0% success rate (pretrained + SFT) vs. 17.8% (Wan2.2 from scratch) - a 31.2% gap
- Real-world Peg Cylinder: 70% success rate / 96% process score vs. π0.5 (50%/90%) and FastWAM (30%/77%)
- Real-world long-horizon tasks: 70–80% success rates (Organize Plate, Arrange Fruits, Cup Stacking) vs. FastWAM at 20–40%
Applied Context: Builders should treat mobile manipulation not as a scaling problem but an alignment problem - decoupling navigation and manipulation into separate prediction heads and training on self-generated video rollouts (Dream Forcing) are both critical for long-horizon reliability. The latent action abstraction also enables transfer across robot embodiments without retraining.