OpenWAM Documentation¶

Heng Yu^*, David D. Yuan^*, Juze Zhang^*, Changan Chen, Yao Feng,
Michelle Baldonado, Steve Cousins, Li Fei-Fei, Jiajun Wu, Ehsan Adeli

Stanford University
^* Equal contribution

OpenWAM is an extensible and composable library for training and evaluating world action models while keeping the shared visual backbone stable. Start with a CPU example, then choose a training, evaluation, or extension workflow.

Start Here¶

• Quickstart: install, validate configs, and run CPU-safe smoke checks.
• Architecture: the stable runtime boundary and core abstractions.
• Compatibility: supported platforms, SDK surface, and numerical gates.
• Policy Architectures And Programs: how topology, conditioning programs, and decoders compose.
• Benchmarks and Data: LIBERO, RoboTwin, CALVIN, and synthetic fixtures.
• Training and Inference: maintained train, resume, eval, sanity, and rollout commands.

Extension Points¶

• Extension SDK: registry surfaces for datasets, policy variants, and decoders.
• Cookbooks: concrete recipes for adding new research components.
• Artifacts: checkpoint manifests, local path aliases, and artifact cards.
• Reproducibility: result envelopes, experiment cards, and tracking policy.

Reference¶

• Research paper (PDF): framework, methods, and experimental results.
• CLI Reference: commands and examples.
• Rollout Contracts: sessions, executed history, and control integration.
• Testing: local tests and numerical comparisons for extensions.
• Releases: package versions and reproducible installation.