HumanWAM-Goal-Base
a base robotics model trained on 60,000 hours of video
2 min readupdated
HumanWAM-Goal-Base is a foundation checkpoint from the Open-HumanWAM project by DAGroup-PKU researchers, built for bimanual robot manipulation. The model acts as a joint generator of future video frames and a 48-step action chunk: it takes a text instruction, a goal image, current feeds from head and wrist cameras with 0.5-second and 2-second history, and the robot's state. The checkpoint was trained on a 60,000-hour hybrid dataset (46,000 hours of first-person human action videos and 14,000 hours of multi-robot data). Its main standout feature is zero-shot task execution (running without task-specific fine-tuning) via a unified end-effector coordinate space (shared-EEF), while also serving as a solid base for adapting to new robots.
Key features
- Joint video and action prediction: simultaneously generates a visual rollout and a 48-step trajectory from a text prompt, goal frame, and multi-camera history.
- Zero-shot control without fine-tuning: executes commands in a unified head-camera NWU coordinate frame for 6-DoF bimanual manipulators.
- Ready for post-training: ships with normalization calibration bounds (
robot.json) for fast adaptation to specific simulators and physical platforms viagoalwam_train. - Fast sampling and standard packaging: weights are exported as a single
safetensorsbundle (including abf16transformer,vae, and tokenizer) and use a 10-step UniPC sampler.
What you can do
- Run zero-shot robot control: test manipulation tasks guided by a text command and a goal photo with no prior fine-tuning.
- Fine-tune for your own robotics platform: use the checkpoint as initialization weights in
goalwam_trainto adapt it to a specific manipulator or simulator like RoboTwin 2.0. - Research human-to-robot skill transfer: use the model as a baseline for experiments with world-action architectures trained on egocentric human videos.
Example
from goalwam import GoalWAM
with GoalWAM("HumanWAM-Goal-Base", backend="native") as model:
result = model.predict(observation)
Read more
DetailsMetrics: 13.94% average zero-shot success rate on the RoboTwin 2.0 benchmark and 30.5% across 20 real-world physical tasks without target fine-tuning.
Training: 60,000 hours of hybrid data (46k hours of egocentric human videos + 14k hours of robot data), saved using EMA weights at step 168,000.
Configuration:
bf16 transformer, UniPC sampler (10 steps, flow shift 2.0), internal gripper calibration gap_m (0.10 m opening, 0.12 m reference width).









Comments