Recurrent-Depth VLA: Implicit Test-Time Compute Scaling of Vision-Language-Action Models via Latent Iterative Reasoning

Benchmark Model Rank Results
zero-shot-generalization-on-calvinRD-VLA (0.5B)#11Avg. sequence length: 3.39
robot-manipulation-on-libero-objectRD-VLA (Fixed)#58Success Rate: 99.0
robot-manipulation-on-libero-goalRD-VLA (Adaptive)#61Success Rate: 96.8
robot-manipulation-on-libero-longRD-VLA (Adaptive)#90Success Rate: 85.8
robot-manipulation-on-libero-spatialRD-VLA (Fixed)#93Success Rate: 92.0
robot-manipulation-on-liberoRD-VLA (Fixed)#118Success Rate: 93.0
robot-manipulation-on-calvinRD-VLA–avg. sequence length (D to D): 3.39