EmbodiedMidtrain: Bridging the Gap between Vision-Language Models and Vision-Language-Action Models via Mid-training

Open paper
Benchmark Model Rank Results
robot-manipulation-on-libero-longInternVL3.5-1B + EmbodiedMidtrain–Success Rate: 54.2
robot-manipulation-on-simplerenv-widow-xInternVL3.5-1B + EmbodiedMidtrain–Average: 0.563
zero-shot-generalization-on-calvinInternVL3.5-1B + EmbodiedMidtrain–Avg. sequence length: 3.714

Rank counts only results with a code link.