If you bridge recorded trajectories with LVLM, then cameras are necessary visual input for LLM to decide which sub-tasks need to be performed to accomplish long

If you bridge recorded trajectories with LVLM, then cameras are necessary visual input for LLM to decide which sub-tasks need to be performed to accomplish long-horizon task, and sub-tasks correspond to pre-recorded ("blind") trajectories which are replayed. If you go beyond pre-recorded "blind" trajectories into more robust task-policies (which you would have to train from many demonstrations) then cameras become necessary to execute the sub-task.