Found first: a primary source the press has not covered yet.
Researchers from PKU, NUS, MIT, UCSB, and NVIDIA have released Open-HumanWAM, a fully open recipe for scaling egocentric human-video pretraining for robot manipulation, with an empirical log-linear scaling law relating pretraining data volume to downstream robot success rates. The paper, submitted November 4, 2026, accompanies the release of Open-HumanWAM-50K: 50,000 hours of annotated egocentric video along with the full curation pipeline, pretrained checkpoints, and evaluation suites.
What the source says
The team fits the relationship SR = a + b log₁₀H with R² between 0.877 and 0.969 across configurations, meaning each order-of-magnitude increase in pretraining hours yields a fixed additive gain in success rate. Using inverse-dynamics attention, scaling from 100 to 50,000 hours raises scene-OOD success from 49.1% to 73.2% and random-configuration success from 21.5% to 42.9%. The resulting world-action models reach 92.08% on RoboTwin 2.0 Clean2Clean and 68.88% on Clean2Random for HumanWAM-Vanilla, trained on human data alone. HumanWAM-Goal, which trains jointly on human and robot data, reaches 84.60% on Clean2Random. The corpus is densely annotated with hand motion and natural language labels, and the full release covers data curation and pretraining pipelines, pretrained and post-trained checkpoints, inference code, and simulation evaluation suites.
Why it matters
Egocentric human video has been used in robotics pretraining before, but no prior work established a quantified, reproducible scaling law or released the full data at this scale. With the fit law public and the corpus available, labs can estimate how much additional pretraining data they need to reach a target success rate without running their own ablations from scratch. The full open release makes the scaling claims independently replicable: anyone can verify the numbers using the same corpus, pipeline, and checkpoints. That lowers the barrier to entry for empirical robotics pretraining research considerably.