UniVLA: Learning to Act Anywhere with Task-centric Latent Actions

Benchmark Model Rank Results
zero-shot-generalization-on-calvinUniVLA (Task-centric Latent Actions, CALVIN ABC→D)#9Avg. sequence length: 3.8
robot-manipulation-on-calvinUniVLA#13avg. sequence length (D to D): 3.80
robot-manipulation-on-simplerenv-widow-xUniVLA (Task-centric Latent Actions)#43Average: 0.479
robot-manipulation-on-libero-goalUniVLA (Task-centric Latent Actions)#99Success Rate: 95.6
robot-manipulation-on-libero-longUniVLA (Task-centric Latent Actions)#101Success Rate: 92
robot-manipulation-on-libero-spatialUniVLA (Task-centric Latent Actions)#104Success Rate: 96.5
robot-manipulation-on-libero-objectUniVLA (Task-centric Latent Actions)#108Success Rate: 96.8
robot-manipulation-on-liberoUniVLA (Task-centric Latent Actions)#131Success Rate: 95.2