arXiv：cs.LG· Timi Owolabi, Hikmah Olawore, Frances Adelakun, Ayotomiwa Oyewumi·· 13 小时前
SmolWAM：面向边缘设备机器人操作的潜在世界-动作模型
SmolWAM: A Latent World-Action Model for Robotic Manipulation on Edge Devices
AI 导读
SmolWAM 是从 2B 参数 Cosmos Policy 派生的潜在世界-动作模型，保留预训练主干并重建推理接口，仅用 12/28 个 Transformer 块和 1,008 个 token，训练耗时 3.6 GPU-hours。
正文
Abstract:World-action models (WAMs) predict a robot's actions together with the observations those actions will produce. However, the strongest models rely on video diffusion transformers that are far too computationally expensive for resource-constrained edge devices. We present SmolWAM, a latent WAM that we derive from the 2B-parameter Cosmos Policy by keeping its pretrained trunk and rebuilding its inference interface: single-frame visual latents, cached text features that preserve the trunk's attention sinks, 12 of 28 transformer blocks, and 1,008 tokens from which actions and future visual latents are read out in one rectified-flow step. Trained in 3.6 GPU-hours, SmolWAM succeeds in 95.5% of 400 LIBERO episodes at 57 ms and 2.3 TFLOP per 16-action chunk with 2.4 GiB of GPU memory, 11x faster than its teacher with 18.5x fewer FLOPs and 3.2x less memory, and its world head still forecasts scene motion.
| Comments: | Accepted at the NeurIPS 2026 Robot Learning with World Models Workshop. Code: this https URL Project: this https URL |
| Subjects: | Robotics (cs.RO); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.13740 [cs.RO] |
| (or arXiv:2610.13740v1 [cs.RO] for this version) | |
| https://doi.org/10.48550/arXiv.2610.13740 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Timi Owolabi [view email]
[v1]
Fri, 9 Oct 2026 15:34:33 UTC (2,069 KB)
来源：arXiv：cs.LG · arxiv.org