Faster-WAM: Do World Action Models Need Deep Action Modules?
Faster-WAM: Do World Action Models Need Deep Action Modules?
This research introduces Faster-WAM, a world-model-centric approach to action prediction that replaces deep action modules with a lightweight expert. The model achieves competitive control performance on LIBERO and RoboTwin~2.0 without additional embodied pretraining, with inference speedups of 3.7× and 1.3× over Fast-WAM and π₀.₅, respectively. It also shows stronger generalization under distribution shifts and comparable real-robot success rates with lower latency and faster task completion.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
inference speedups
3.7
inference speedups
1.3
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
oTwin~2.0 without additional embodied pretraining. It provides approximately $3.7\times$ and $1.3\times$ inference speedups over Fast-WAM and $\pi_{0.5}$, respectively. Consistent with its world-model-centric design, Faster-WAM demonstrates stronger generalizability under distribution shifts: the same LIBERO-trained policy achieves $78.3\%$ success on LIBERO-Plus, exceeding Fast-WAM and LingBot-VA by $26.8$ and $8.8$
Open source S5
percentage points, respectively. Finally, real-robot experiments demonstrate success rates comparable to Fast-WAM, with substantially lower inference latency and shorter task-completion times.
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org