HumanWAM-Vanilla-Base
HumanWAM-Vanilla base checkpoints for robot post-training. This repository contains the checkpoint pretrained on 50,000 hours of Open-HumanScale-50K, together with its EgoPro multi-view mid-trained continuation.
Checkpoints
| File | Stage | Training step | Initialization |
|---|---|---|---|
pretrained/step_040000.pt |
50,000-hour human egocentric pretraining | 40,000 | Human pretraining stage |
midtrained/step_010000.pt |
EgoPro multi-view adaptation, W0 setting | 10,000 | The pretrained checkpoint above |
Step numbers are local to each training stage. The mid-trained checkpoint is the initialization used for the released HumanWAM-Vanilla RoboTwin policy.
Release status and use
This is a weights-first release. Normalization statistics, matching configurations, and inference resources will be added separately. These PyTorch checkpoints require the matching HumanWAM implementation and stage-specific resources; the files alone are not a complete runnable package.
The base checkpoints are intended for downstream robot adaptation. Use the task-specific RoboTwin repository for the selected RoboTwin policy. Action transforms and normalization statistics must match the training stage and downstream embodiment.
File integrity
| File | Bytes | SHA-256 |
|---|---|---|
pretrained/step_040000.pt |
12041838137 | e0a00e55483d8c08376be2650f159da1d2ce4ec22d5cfcf925a28a310de6cbf6 |
midtrained/step_010000.pt |
12041838009 | c87cee0ee7fac06ed55f809230e2c83c26f69b4c3239ce561e191df8aaccaf9f |