Why Do VLA Robots Tremble? Diagnosing Temporal Discontinuity in Vision-Language-Action Models

Vision-language-action (VLA) models hold the promise of general-purpose robot control through large-scale vision-language pretraining. However, we reveal that current VLAs produce structurally jerky trajectories even on their own training distribution---a deficiency invisible to standard task success and reconstruction error metrics. Through systematic temporal analysis of three architecturally distinct VLA families (OpenVLA, Octo, π₀), we trace the root cause to a paradigm mismatch: VLAs inherit per-image independent processing from their vision-language backbones, but robot actions demand physical continuity between consecutive timesteps. We derive five falsifiable predictions from this diagnosis and verify four: (P1) continuous action heads exhibit the same discontinuity as discrete ones, (P2) temporal averaging reduces jerk as N^(-0.87) (R² > 0.99), (P3) chunk-based models show positive autocorrelation within chunks but negative autocorrelation at chunk boundaries (p < 10⁻⁵, Cohen's d = 1.51), and (P5) noise-induced jerk scales as f^3.1 with control frequency. Physical validation on an xArm Lite 6 robot confirms signal fidelity: commanded jerk 4.9× above ground truth is faithfully executed without triggering any protective mechanism---a silent compliance that underscores deployment risk.