Parameter-efficient adaptation of tri-view robot video predictors

Runs and adapts a released tri-view predictor on limited hardware; local error reduction still trails static persistence, and the historical scored system is distinct from the adapter.

Earlier version, not an independent replication or a new official score for a later adapter.

Why we tried it

Adapt a released tri-view predictor on constrained hardware and test whether gains depend on actions.

What the evidence shows

After 96 updates, local adaptation reduced matched initial-model error by 71.06% on ten explored trajectories, but still trailed first-frame persistence.

Where the work stands

The historical scored system, adapter and prototype are distinct; ten reused trajectories do not establish causal action contribution or unseen-task validity.

Possible next steps

Use untouched task groups and paired action-divergent branches, vary adapter seeds, and bind exact model and evaluation artifacts.

Suggested next steps are untested research questions, not new experimental results. The archive records manuscript provenance, licensing and status.