Pretrain, then fine-tune
240 MOTION ROWS+ 16 ACTION ROWS
observed state transition → reused policy weights → scarce actions
—HELD-OUT ACTION MSE
Waiting for real training.
Same policy. Different memory.
Measure whether motion-derived weights help scarce action learning. This synthetic result is not robot, company, deployment, safety, or investment evidence.
+ 16 ACTION ROWS
Same model, from scratch
0 MOTION ROWS+ 16 ACTION ROWS
random weights → same scarce actions → same held-out test
—HELD-OUT ACTION MSE
Waiting for the fair baseline.
Run the shared test.
Both policies will use the same architecture, action rows, and 80 held-out states. Initialization history is the intended difference.
0 / 80 TEST PREDICTIONS
What would actually prove it?
Real multimodal perception. Unseen-task transfer. Closed-loop hardware control. Latency, safety, recovery, and deployment evidence.
Hold architecture constant
A transfer claim needs a baseline with comparable capacity, optimization, downstream rows, and one untouched test set.
Test shift, not just fit
Useful representations must survive changed objects, lighting, embodiment, dynamics, and task instructions.
Separate research from valuation
A synthetic loss curve cannot establish product readiness, deployment traction, or public and private company value.