Coda Reduces VLA Latency via Learned Endpoint Correction

Coda optimizes flow-matching VLA policies by replacing high-step integration with a few-step trajectory followed by a lightweight Transformer-based residual corrector. On RoboTwin Easy tasks, a two-step configuration achieved a 2.12x speedup while maintaining competitive success rates.