Dyna

Customer orientation: quality, throughput and reliability the customer site requires.

Customer orientation: Judges real performance based on real on-site testing. Pass rates are drawn from on-site reporting, graded against customer acceptance criteria, by operators not involved in model development. Both models are post-trained on the same task datasets for the same number of steps, and neither has seen any data from the deployment sites.

VLA vs WAM: VLA training samples short, sparse windows, so it pays that seek penalty on every sample. World-action models read long contiguous sequences, which amortize one keyframe over many frames, so we can afford larger GOPs.

Storage method limits and cross-embodiment scaling conflicts solved:

Q: But in the scaling dyna 2 article it says that cross embodiment scaling emphasizes scaling of video in pre training, but in infra dyna-2 article it says that training needs a lot of action rather than video. Does that mean that scaling video will become hard due to need of large action volume?

A: Looking at the other article (cmswq68jt0000l704xlyyrw5t) carefully — cross-embodiment in the Dyna-2 scaling paper includes both human egocentric video and data from multiple robot embodiments, not just human footage. So the corpus is a mix: human video (no action stream needed), plus robot episodes from various embodiments (action streams required). This means the tension you originally identified is real but partial — it applies to the robot portion of the corpus, not the human video portion. Scaling the robot side does require paired action sequences, but the human video side can scale freely without them. The 1 million hours figure cited in this infrastructure article (¶) refers to egocentric video broadly, which would include both. So the practical constraint is: human video scales relatively freely, robot episode scaling is gated on action stream quality and availability, and the ingestion infrastructure (¶) is what makes the robot side tractable at this volume.

It is why we built the foundation instead of scripting our way through: researchers can now ingest, curate, and experiment with orders of magnitude larger data without rebuilding the path each time, which accelerates our iteration cycles and widens the range of experiments worth designing.

The aggregate results over 7 benchmark tasks are shown below. Pooled over all trials and checkpoints, the WAM reaches 1.55× the VLA's success rate and 1.12× its grade. In head-to-head comparisons — fixing both the pre-trained checkpoint step and the post-training task — the early Dyna-2 won 65% of the time, Dyna-1 won 29%, with 6% tied. Most of the VLA's wins come at the earliest pre-training checkpoint, before the WAM model's pre-training advantage had accumulated.