Open Instruction Support

[zhuqiangLu]

Just wondering does JarvisVLA support open instruction?

According to the evaluation scripts, it seems it only support predefined tasks.

[MuyaoLi-jimo]

Just wondering does JarvisVLA support open instruction?

According to the evaluation scripts, it seems it only support predefined tasks.

You’re right — Jarvis-VLA only supports a predefined instruction set. In this work, our primary focus is to investigate how world knowledge influences the control policy. Extending it to an open-ended instruction space is feasible. A commonly used approach is to leverage internet-scale trajectories (e.g., from YouTube) to construct expert datasets, as demonstrated in Minedojo,