A multi-university OpenWAM team submitted OpenWAM-α on Sept. 7, 2026, an open modular world-action pretraining stack with code and weights. That matters because world-action systems are often closed and tightly coupled; a factorized open recipe would let outside labs swap backbones and re-run which design choices actually move control quality, if author-reported sim and real gains hold.
In plain terms, OpenWAM-α is not a chat model and not a finished factory robot product. It is a world-action model: a video-generation backbone that jointly predicts future scenes and robot actions so the same prior can drive arms and hands across simulation and real hardware, rather than a vision-language policy bolted on after the fact.
The technical report (arXiv:2609.07398) splits the work into OpenWAM-Infra (pluggable encoders and attention families plus an eight-sim harness), OpenWAM-Study (three design principles from controlled ablations), and OpenWAM-α (Wan2.2-TI2V-5B video backbone, Dual-System Mutual attention, about 1.2 billion action parameters, 80-D unified action space). Authors say pretraining used 518.5 million frames, about 6,400 hours, mixing roughly 70% robot data with 30% egocentric human video. Treat the corpus size and recipe as author claims.
On Franka Research-3 single-arm trials, authors report OpenWAM-α at 99 of 120 successes (82.5%) versus LingBot-VA at 77.5% and pi0.5 at 55.0%. On RoboDojo real bimanual (ARX X5, Piper, Piper X), they report overall average Score/SR of 37.6 / 24.4 and say that tops their leaderboard versus X-VLA and Xiaomi-Robotics-0. On an unseen Wuji hand plus Tianji arm embodiment, they report higher in-domain and out-of-domain success than pi0.5, including Put Away Clothes in-domain at 100% versus 60%. Authors also flag LIBERO-Plus as a weak spot tied to thin single-arm pretrain coverage. Those figures are author-protocol claims, not third-party re-runs.
AlphaXiv and an arXiv TLDR page independently restate the Wan2.2-plus-Mutual recipe and the 518.5 million-frame mix without a controlled re-evaluation. Full stack assets sit at openwam-official.github.io, GitHub OpenWAM-Official/OpenWAM (Apache-2.0, about 820 stars as of Sept. 21, 2026), and Hugging Face org OpenWAM with Alpha and Study checkpoints. The pack is distinct from peer Discover world-model stories Odyssey-3, XPACE, UnifoLM, Pelican-Sim, Fire3D, LingBot, WorldSculpt, and Dream-RSI.
What remains open is whether third parties can match the Franka, RoboDojo, and dexterous success rates under shared protocols, how far LIBERO-Plus underperformance limits the top-tier claim, and whether the Study principles hold when other labs swap backbones. Author benches and outsider summary restatements should be weighed separately until those points are settled.