We benchmarked 7 hours of Human Archive data against four widely used public datasets totalling 17 hours: HOT3D, H2O, HOI4D and TACO. We ablated only the pre-training data mixture, holding the model, the fine-tuning procedure and the evaluation identical across every condition.
Results
Task success
Share of evaluation episodes completed end to end.
Human Archive only · 7 hrs85.0%
4 public datasets · 17 hrs50.0%
Mixture of all 5 · 24 hrs45.0%
0%50%100%
Task progress
Fraction of each task's stages reached before failure.
Human Archive only · 7 hrs90.6%
4 public datasets · 17 hrs76.3%
Mixture of all 5 · 24 hrs78.8%
0%50%100%
Setup
Policy
UniDex VLAVision-language-action model, Tsinghua University
Robot
Franka arm with ProHand 1.0Dexterous hand by Proception AI
Human Archive data
7 hoursEgocentric stereo video with 3D hand pose and world-frame SLAM
Public baselines
HOT3D, H2O, HOI4D, TACO17 hours combined
Variable
Pre-training data mixtureThe only thing that changed between conditions
Held constant
Model, fine-tuning, evaluationIdentical across all three runs
Method
Three conditions. Pre-train on Human Archive data alone, on the four public datasets alone, or on all five combined.
One variable. The pre-training corpus. Model architecture, fine-tuning procedure and the full evaluation protocol were identical in every run.
Two metrics. Task success counts completed episodes; task progress credits partial completion, so a policy that reaches the final stage and fails still scores above one that never grasps the object.