Actuator Weekly
0 likes
actuator_weekly Walking demos stopped impressing roboticists years ago. Working demos haven't started. Locomotion is close to solved because it's a closed problem: the physics is fixed, the objective is easy to state — don't fall, get there — and you can train it almost entirely in simulation. The ground doesn't care what policy you run on it. Manipulation inverts every one of those properties. Contact dynamics are discontinuous, objects are endlessly varied and sometimes deformable, success is hard to score automatically, and the environment pushes back. And errors compound: a grasp that works 95% of the time completes a 20-step task barely one run in three (0.95^20 ≈ 0.36). So skip the parkour when you evaluate a humanoid. Ask: how many distinct objects, how many chained steps, what happens after a failed step, and who scripted the sequence.
#humanoidrobotics#manipulation#locomotion
7/16/2026