Open-HumanWAM
Collection
Model checkpoints for Open-HumanWAM. • 6 items • Updated • 5
Pretrained HumanWAM-Goal foundation checkpoint from Open-HumanWAM (hybrid human–robot pretraining, EMA weights).
HumanWAM-Goal jointly denoises future video and a 48-step action chunk from a text instruction, one goal image, current head and wrist camera observations with a 0.5 s / 2 s visual history, and the robot state. This checkpoint supports:
domain_id = -1, bimanual 6-DoF end-effector poses in
the head-camera NWU frame, achieving 13.94% zero-shot mean success on RoboTwin 2.0 and
30.5% across 20 real-world tasks without target-task fine-tuning).goalwam_train), carrying the shared EEF
normalization bounds in robot.json used by downstream post-training (such as
DAGroup-PKU/HumanWAM-Goal-RoboTwin).| Item | Value |
|---|---|
| Stage | Hybrid human–robot pretrained weights (EMA) |
| Format | Safetensors export: transformer/ (bf16 safetensors), vae/, text_tokenizer/, processor.json, sampling.json, robot.json, export.json |
| Shared EEF convention | Head-camera NWU poses (26:33, 34:41), gap_m internal gripper calibration (0.10 m open aperture, 0.12 m reference width) |
| Sampling | UniPC, 10 steps, flow shift 2.0 |
hf download DAGroup-PKU/HumanWAM-Goal-Base --local-dir HumanWAM-Goal-Base
from goalwam import GoalWAM
with GoalWAM("HumanWAM-Goal-Base", backend="native") as model:
result = model.predict(observation)
clean_50)
torchrun --nproc-per-node 8 -m goalwam_train.train \
--pretrained HumanWAM-Goal-Base \
--data /path/to/robotwin_clean_50 \
--goal-cache /path/to/train_goal_cache \
--output /path/to/outputs
See the post-training documentation for environment setup, data conversion, normalization, and evaluation.
| File | Bytes | SHA-256 |
|---|---|---|
export.json |
2639 | c17dc6e933a8ec58cb43f8fe2b7dd69f667afd599bc4b90133448f085a59577c |
robot.json |
5317 | 80c733d1e4d17f79c1088e47a2761d28c739648a58434e6451943786a1d92783 |
All checkpoint files are listed with their SHA-256 in export.json.