MotionWAM

Whole-Body Loco-Manipulation

Highlights

1. Unified whole-body motion latent. A single action space of whole-body motion tokens that jointly covers locomotion, torso motion, height regulation, foot interaction, and hand manipulation, replacing the upper–lower decoupled interface of hierarchical humanoid systems and unlocking task-driven foot behaviors.
2. Three-stage egocentric-to-embodiment training. Egocentric video pretraining on 2,136 hours of human and humanoid video, cross-embodiment action post-training, and whole-body teleoperation fine-tuning — progressively adapting a video world model to first-person dynamics and the target humanoid embodiment.
3. State-of-the-art real-world performance. On nine Unitree G1 whole-body tasks, MotionWAM lifts the overall success rate from $43.9\%$ (GR00T-N1.7) to $\mathbf{76.1\%}$, an over +32% absolute gain over the strongest VLA baseline trained on the same demonstrations.
4. Real-time WAM for humanoid loco-manipulation. An end-to-end World Action Model that conditions the policy on intermediate denoising features of a video world model, running closed-loop at 4.9 Hz on a single A100 — 7× faster than comparable world-model-based policies such as Cosmos Policy.

Method

MotionWAM is a dual-DiT World Action Model trained in three stages. Stage 1: the Video DiT is pre-trained alone on egocentric human and humanoid videos. Stage 2: the Motion DiT is attached and co-trained across heterogeneous Unitree G1 datasets via specific embodiment tags, conditioned on Video DiT hidden states to predict discrete motion-token index and continuous end-effector values. Stage 3: the full model is finetuned on teleoperated whole-body demonstrations retargeted from SMPL-24 to Unitree G1.

Real-World Results

We evaluate MotionWAM on a suite of nine real-world whole-body loco-manipulation tasks on the Unitree G1 humanoid. All baselines are fine-tuned on the same Stage 3 demonstrations and emit actions through the same SONIC interface.

MotionWAM wins on every task and lifts the overall success rate from $43.9\%$ (GR00T-N1.7) to $76.1\%$, an over $32\%$ absolute gain. The gap is largest on tasks demanding whole-body coordination beyond the upper limbs — Kick Soccer ($+40\%$), Load Cart ($+40\%$), Retrieve Item ($+40\%$), Wipe Board ($+45\%$), and Do Laundry ($+30\%$) — where the unified motion latent gives MotionWAM access to task-driven foot and torso behaviors that a decoupled upper–lower interface cannot express.

Task Demonstrations

Ablation: Three-Stage Training

We disentangle the contribution of each pretraining stage by disabling one stage at a time while keeping Stage 3 fixed. Removing Stage 1 (egocentric video pretraining) costs $11\%$ absolute success; removing Stage 2 (cross-embodiment action post-training) costs $28\%$. The two stages play complementary roles: Stage 1 supplies an egocentric visual-dynamics prior, while Stage 2 grounds that prior into the action space across embodiments.

Real-Time Inference

Because MotionWAM conditions on a single forward pass of the Video DiT rather than a full denoising rollout, it remains real-time on a closed-loop humanoid. On an NVIDIA A100, MotionWAM runs at $4.9$ Hz at the chunk-wise rate — competitive with VLA baselines of similar scale, and seven times faster than Cosmos Policy, the closest world-model-based policy at comparable parameter count.

BibTeX

@misc{zheng2026motionwamfoundationworldaction,
      title={MotionWAM: Towards Foundation World Action Models for Real-Time Humanoid Loco-Manipulation}, 
      author={Jia Zheng and Teli Ma and Yudong Fan and Zifan Wang and Shuo Yang and Junwei Liang},
      year={2026},
      eprint={2606.09215},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2606.09215}, 
      }