A Beginner's Guide to Humanoid Robot Hands and Manipulation

Walking gets the highlight reels. Picking things up reliably is arguably the harder unsolved problem.

A humanoid that can walk gracefully but can't reliably pick up an unfamiliar object is far less useful than the walking demo suggests. Manipulation — and specifically grasping things the robot hasn't seen before — remains one of the field's genuinely open problems.

Hand design tradeoffs

Figure's fourth-generation hand reportedly packs 16 degrees of freedom per hand with sensors in each finger, aimed at fine force control for tasks like handling small components without crushing them. Unitree's Dex3-1, by contrast, uses a simpler three-finger design with around 33 tactile sensors across the fingers and palm.

More degrees of freedom generally means finer control, at the cost of more actuators, more mass in the hand itself, and a harder mechanical design problem — cable and tendon routing through a multi-DoF wrist without binding or tangling over thousands of cycles is a real, unglamorous engineering challenge.

Why grasping is still hard

Grasp planning for deformable or non-rigid objects — bags, cables, cloth — remains genuinely unsolved, because rigid-body grasp models don't capture how the object will actually behave once contact starts.

A lot of the manipulation shown in production demos still leans heavily on teleoperation, particularly for anything involving fine force control or a genuinely novel object. Autonomous grasp success rates on unstructured, previously unseen clutter are still well below what a human teleoperator can achieve.

What's actually improving

Tactile sensing has come a long way for coarse contact detection across a large area — knowing something is touching the hand and roughly how hard. Fine, high-resolution force distribution sensing, closer to what a human fingertip provides, is still the harder half of the problem, and it's where a lot of current research effort is concentrated.