MiVLA on LocoMan

The complete real-robot pipeline for a quadruped dual-arm platform.

MiVLA is a generalizable vision-language-action model trained through human-robot mutual imitation. It aligns human and robot action spaces and learns from simulated robot demonstrations and human videos without real-robot data during pre-training.

My role

I independently delivered the complete LocoMan real-robot workflow:

  • Installed, integrated, and debugged the quadruped dual-arm platform
  • Built a VR teleoperation data-collection system based on Human2LocoMan
  • Integrated the platform with Isaac Gym for simulation validation
  • Deployed and validated trained policies on the physical robot
  • Conducted all real-robot experiments reported for the LocoMan platform

Result

MiVLA outperformed leading VLA baselines by 25% in simulation and 14% on real robots. On the composite LocoMan embodiment, mutual imitation was particularly important: the complete bidirectional objective achieved a 50% success rate, compared with 20% for the one-way variant.

Technologies: VLA, LocoMan, Human2LocoMan, VR teleoperation, Isaac Gym, PyTorch

Read the paper