TDSR-VLA: Transition-aware Denoising Sequence Representations for Vision-Language-Action

Benchmark Model Rank Results
robot-manipulation-on-libero-longTDSR-VLA#108Success Rate: 56.0