OpenWAM: NUS-led open full-stack world-action model tops real-robot leaderboard with 6,400h pretraining
青稞AI · wechat · 2026-10-12
Researchers from NUS, Peking University and others released OpenWAM, an open full-stack that turns world-action model pretraining into controlled, modular experiments.
The stack comprises three parts: OpenWAM-Infra decomposes the WAM design space into composable modules with unified training, inference, deployment and evaluation across 8 simulation benchmarks; OpenWAM-Study distills three design principles from controlled ablations (knowledge transfer via strong generative backbones and compact latent spaces; dedicated action capacity, explicit world-to-action information flow and synchronized joint denoising; embodied pretraining mainly boosting out-of-domain generalization); and OpenWAM-α, pretrained on 6,400 hours (518.5M frames) of egocentric human and robot data, achieves SOTA on the EBench mobile bimanual benchmark and ranks first on the real-world RoboDojo-Real leaderboard, doubling pi0.5's success rate. Infrastructure, evaluation protocols, pretrained models and data recipes are all open-sourced.
More from Embodied
- Robotic systems and real-time connectivity let surgeons operate on patients thousands of km away — CurieuxExplorer · 2026-10-12
- Humanoid robots learn Nadal and Federer's tennis styles straight from broadcast footage — CurieuxExplorer · 2026-10-12
- Humanoid hands are the hardest problem: Shadow, Mimic and PSYONIC specialize in them — xmercury_one · 2026-10-12
- Evolved e-puck robots show communication emerges with information value — uv-mex · 2026-10-12
- Atomic Machines exits stealth with Matter Compiler, an AI-native manufacturing system — mikeflache · 2026-10-12
- Meta debuts 'I Let You In,' a Talk to Me universe VR horror series for Quest — ZeroStateReflex · 2026-10-12