Researchers find simple fix for random seed failures in VLA robot training

Fine-tuning a vision-language-action model (VLA-JEPA) on a single GPU should be straightforward, but researchers at Skoltech have uncovered a hidden trap. Running the same training script 13 times with different random seeds, they found that 12 runs produced robots succeeding 91–94% of the time — while one run silently degraded to 65.2%, a 29 percentage point gap with no error message. They term this the 'seed lottery' and trace its cause to output collapse: the action predictor learns to produce nearly identical outputs regardless of visual input. Existing weight-level regularization (L2, EWC) structurally misses this collapse because it penalizes weight changes, but collapse occurs in directions weights can move freely — a gap formalized via the Jacobian null-space.

Across 7 methods × up to 13 seeds × 3 LIBERO benchmarks, the team found that three output-level regularizers — VICReg (12 seeds tested), Dropout (4 seeds), and a halved learning rate (5 seeds) — each eliminated every catastrophic seed. In total, 0 out of 21 combined collapses versus 1 out of 13 for the baseline (F(12,11)=28.7, p<0.001). Weight-level methods preserved the lottery. The simplest fix is changing one number in your optimizer config. This research, submitted to CoRL 2026, provides a practical, zero-cost solution to a silent failure mode in robot model deployment.