機器人與世界模型
DECOWAM Separates Robot Dog Base and Arm Representations, Cutting Action MSE by 21.7% While Updating Just 25.95 Million Parameters
DECOWAM adds base, arm, and camera egomotion interfaces to the frozen 6.02-billion-parameter FastWAM, using only the current observation at deployment to predict 48 action steps. Its overall completion rate in physical robot tests is only slightly higher than FastWAM’s, but its robustness success rate under base disturbances rises from 12.7% to 30.4%.
World-action models for fixed robotic arms typically assume a static camera viewpoint. When the camera is mounted on a quadruped base, however, image displacement simultaneously mixes object motion, arm actions, and body movement. Rather than packing the 14-dimensional whole-body action into a single latent representation, DECOWAM creates separate representations for the base and arm, uses gradient reversal to reduce information leakage between them, and injects the 3D base velocity directly into the video tokens to account for camera egomotion. The [paper](https://arxiv.org/abs/2608.20114) also uses a teacher with access to future frames to compress an “equivalent future” into a student bottleneck that reads only the current frame and robot state. The teacher and future inputs are removed at deployment, so the resulting policy remains causal.
The system uses the [open-source FastWAM implementation](https://github.com/yuantianyuan01/FastWAM) as its backbone. In the second stage, it freezes approximately 6.021 billion parameters and trains only 25.95 million adapter, bottleneck, and conditioning parameters—a roughly 232-fold reduction. Across a fixed set of 23 replay sequences, frame MSE falls by 15.0% and 48-step action MSE by 21.7%, at the cost of an 11.4% increase in evaluation latency.
In the more consequential 79 physical robot trials, DECOWAM completes 46 tasks for a 58.2% completion rate, only slightly above FastWAM’s 57.0%, but reduces average completion time by 16 seconds. Its whole-body coordination rate rises from 34.2% to 44.3%, while its success rate under base-displacement disturbances increases from 12.7% to 30.4%.
Data remains the main limitation. Although ARMDOG contains 1,487 synchronized sequences totaling 343,550 frames, 56% of the data focuses on bottle pick-and-place tasks, while the replay evaluation covers only eight box-related tasks. The paper also says only that the dataset and conversion tools are “planned for release.” Engineers should watch whether the complete dataset, model weights, and safety controls are made public, and whether the reported advantages can be reproduced across robot embodiments, environments, and larger-scale physical trials.