MotionWAM: Towards Foundation World Action Models for Real-Time Humanoid Loco-Manipulation

Summary

MotionWAM is a real-time World Action Model for Unitree G1 humanoid loco-manipulation from one egocentric camera. It uses video world-model features and whole-body motion tokens to control legs, torso, height, feet, and hands in one action space.

Problem

Existing WAM policies are too slow for closed-loop humanoid control because they run iterative denoising over video-action latents.
Many humanoid systems split control into upper-body manipulation and lower-body base commands, which blocks task-driven foot actions such as pedal stepping or ball kicking.
The problem matters because real humanoid tasks need coordinated balance, locomotion, torso motion, reaching, and object contact in the same policy loop.

Approach

MotionWAM pairs a Video DiT initialized from Cosmos-Predict2.5-2B with a Motion DiT that predicts whole-body motion latents.
The Video DiT runs one forward pass at a high-noise flow step and exposes intermediate denoising activations; the policy uses those features instead of waiting for a fully generated future video.
The action output is a unified motion latent decoded by SONIC into joint commands, with discrete motion tokens for whole-body movement and continuous channels for grippers or dexterous hands.
Training has 3 stages: pretrain the video branch on about 2,136 hours of egocentric human and humanoid video, post-train video plus action on heterogeneous Unitree G1 data, then fine-tune on 200 teleoperated episodes per task across 9 target tasks.

Results

On 9 real-world Unitree G1 tasks with 20 trials per task, MotionWAM reaches 76.1% average success versus 43.9% for the strongest baseline, GR00T-N1.7, a 32.2 point gain.
MotionWAM reports the largest task gains on Kick Soccer (+40), Load Cart (+40), Retrieve Item (+40), Wipe Board (+45), and Do Laundry (+30) compared with the strongest listed baseline results.
In a 5-task ablation, the full 3-stage model reaches 70.0% average success; removing Stage 1 drops to 59.0%, and removing Stage 2 drops to 42.0%.
MotionWAM runs at 4.9 Hz on one NVIDIA A100 with 2.5B trainable parameters, compared with 0.7 Hz for Cosmos Policy, 6.5 Hz for GR00T-N1.7, and 9.0 Hz for Qwen3DiT.
The paper claims task-driven foot interaction on real hardware, including ball kicking and pedal stepping, which the evaluated decoupled upper-lower control designs cannot express through their action interfaces.

Link

https://arxiv.org/abs/2606.09215v1