What it is
Modular research stack for training, fine-tuning, and evaluating world-action models, with an open pretrained model trained on 518.5M frames. Build, fine-tune, deploy, and compare world-action models with modular configs for architectures, backbones, datasets, and benchmarks, plus released pretrained checkpoints and benchmark adapters.
Key results
- OpenWAM-α is pretrained on 518.5M frames, about 6,400 hours
- Supports RoboTwin, LIBERO, LIBERO-plus, RoboCasa365, VLABench, EBench
- Released checkpoints are self-contained and deploy directly with scripts/deploy.sh
Shows that world-action model pretraining can be opened up as a modular, benchmarked stack rather than a tightly coupled one-off system.