A preprint reports that DECOWAM completed 46 of 79 physical robot trials on one quadruped–arm platform, a 58.2% task-success rate. The paper reports 57.0% for FastWAM and 15.2% for X-WAM; each method was evaluated with common observations, language inputs, low-level control and safety constraints. DECOWAM’s mean completion time was 49 seconds, reported as 16 seconds faster than FastWAM and 33 seconds faster than X-WAM.
In a separate fixed replay test, DECOWAM recorded lower frame and action mean-squared error (MSE), a measure of prediction error, than Original FastWAM at the 50k checkpoint. DECOWAM’s frame MSE was 8.77e-4 versus 1.032e-3 for Original FastWAM, while action MSE was 5.38e-5 versus 6.87e-5. Its peak signal-to-noise ratio (PSNR), an image-quality measure, was 31.663 versus 31.441. The reported relative differences were 15.03% lower frame MSE, 21.71% lower action MSE and 0.222 dB higher PSNR.
The test behind the numbers
The replay used a fixed 23-episode slice from eight box-manipulation task folders and 4,323 frames. Models predicted eight future RGB frames at 384 × 320 pixels and a 48-step, 14-dimensional action chunk.
A model built around the robot
DECOWAM separates base motion, arm manipulation and camera ego-motion—the camera’s own movement—through dedicated conditional interfaces. It also uses residual adapters, an action-equivalent future bottleneck distilled from privileged observations, adversarial base/arm latents and base-velocity conditioning for video prediction.
Future observations and the privileged teacher are used only during training; at deployment, DECOWAM uses the current observation, robot state and language instruction.
Training adapted all FastWAM parameters to ARMDOG for 50,000 steps, then froze them for a Stage-2 path with four structured groups. The reported Stage-2 trainable footprint fell from 6,020.75 million to 25.95 million parameters, an approximately 232-fold reduction. Evaluator latency was reported as 11.4% higher.
Additional tests on the physical platform
DECOWAM also reported 87.3% docking success, 44.3% whole-body coordination success, 30.4% base-displacement robustness success and 32.9% autonomous recovery success. It tied the highest docking rate and led the other three reported robustness measures.
The result depends on the comparison
Among the reported world–action model references, DECOWAM ranked first on every reported video and action metric under a matched eight-frame, 384 × 320 interface. Compared with FastWAM, Motus and X-WAM, its reported frame MSE was lower by 15.03%, 83.1% and 66.6%, respectively; action MSE was lower by 21.71%, 89.3% and 91.5%.
Those comparisons are not fully like-for-like across all references: some WAM outputs used different frame counts or resolutions, which limits direct metric comparisons.
Against action-only vision-language-action systems, which predict actions without the future-video output used by DECOWAM, X-VLA had the lowest error across all three action metrics on the same 23-episode slice. DECOWAM ranked second on A-MSE and A-L2; its reported differences versus π0.5 were 69.9% on A-MSE and 21.1% on A-L2, while its A-MAE was within 11.1% of π0.5 and 84.0% lower than GR00T.
What the evidence does not show
The physical-trial percentages are descriptive, not uncertainty estimates: each method was tested in 79 trials, but the study reports no confidence intervals or significance tests. The replay comparison used a fixed 23-episode slice, and the ablation suite was trained separately. The results do not isolate any individual module’s contribution.
Evidence is limited to one ARMDOG data regime and one physical platform. It does not show that DECOWAM is the best action-only policy—X-VLA had the lowest reported action errors—or that the result transfers to other robots, datasets or environments.
The paper describes a planned release containing raw and cleaned HDF5 files, the world–action conversion, immutable split manifests and a datasheet.
Paper data and sources
Original title: DECOWAM: Decoupled Whole-Body World-Action Model for Legged Mobile Manipulation
Authors: Siyuan Ma, Boshi Zhang, Yutian Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text