XR-1 VLA Temporal Alignment

Testing whether action-prefix conditioning provides latency robustness

I audited the native action-prefix path of the 5.1B-parameter XR-1 VLA and built a causal temporal-counterfactual benchmark using self-generated old-plan prefixes.

Impact

  • 190 formal generations across 4×RTX 4090
  • Measured 288 ms mean / 318 ms p95 generation latency
  • Demonstrated that self-prefix conditioning worsened right-arm handoff error at medium and large phase offsets
  • Established selective post-training feasibility for 675M parameters

This is a real-model systems study, not a physical-robot task-success claim. Its conclusion is a useful boundary: native prefix support alone does not guarantee asynchronous latency robustness.

Code and full evidence