Why I’m reading this

VLA models are usually evaluated on arms with parallel grippers. A tendon-driven hand has compliance, coupling, and friction that a position-controlled gripper does not — and I want to know which of those the policy is implicitly relying on.

What I’m collecting

  • Action spaces. Joint positions, end-effector poses, or latent actions — each one hides a different assumption about the controller.
  • Data. How much of the success is the policy, and how much is a well-tuned impedance controller underneath it?
  • Hardware-aware policies. Work that conditions on hand morphology or tendon state instead of treating the hand as a black box.

Status

Ongoing reading notes. I will publish short summaries as I finish each paper.