ActionPiece: Rethinking Action Tokenization for Autoregressive Vision-Language-Action Models

Benchmark Model Rank Results
robot-manipulation-on-simplerenv-widow-xActionPiece#17Average: 0.719
robot-manipulation-on-libero-plus-zero-shot-transferActionPiece + Qwen3-VL-4B (zero-shot)#41Success Rate: 68.8
robot-manipulation-on-liberoActionPiece + Qwen3-VL-4B#113Success Rate: 94.8