Why walking is easier than picking up a pen: debate on VLA training difficulties

Centered on the question "why is it easier to train a robot to walk than to pick up a pen off the floor," jonstephens85 and Darpinian engaged in a discussion, with carlosdponx later adding technical explanations. The takeaway: picking up a pen looks trivial but the demands it places on gripper hardware and whole-body coordination are underestimated, while VLAs like Pi-0.5 on humanoid robots also face a disconnect between leg and arm control.

Confirmed

• jonstephens85 questioned why VLAs like Pi-0.5 shouldn't be able to automatically identify and pick up a pen, seemingly without extensive post-training.
• Darpinian responded: picking up a pen actually requires almost no visual perception; the real difficulty is that pens vary enormously in shape and appearance, and typical robot grippers are far worse than human hands at picking small objects off a flat surface for a variety of reasons.
• Darpinian further clarified his original point: training pi05 from scratch is itself hard; doing post-training with pi05 to pick up a pen isn't hard, but pi05 cannot control a humanoid robot's legs to keep balance—another system is needed for leg control, so coordinating legs and arms becomes the sticking point.
• carlosdponx offered three technical explanations: actions that don't rely on external perception reduce perceptual uncertainty; the body is a rigidly connected whole, making it easy to simulate; and training such actions is simpler.

Why it matters

• The discussion clears up a public misconception about VLA capability boundaries: whether a model can "pick up a pen" is not just a perception and policy problem, but is also constrained by end-effector hardware and whole-body humanoid coordination (separate leg/arm control).
• For humanoid robotics R&D, integrating VLAs with leg balance control systems is a more critical yet more easily overlooked bottleneck than single-arm manipulation tasks.

2026-10-15 ~ 2026-10-15 · 5 related posts

• Robot grippers, not vision, make picking up a pen hard — Darpinian · 2026-10-15
• Debate: why grippers struggle to pick a pen off the floor — jonstephens85 · 2026-10-15
• [source] Why robots walk easier than they pick pens: a VLA training debate — Darpinian · 2026-10-15
• [source] Pi-0.5 can post-train to grab a pen, but legs-balance coordination is the hard part — Darpinian · 2026-10-15
• [source] Why robots learn to spin before picking up a pen: rigid bodies, self-contained state — carlosdponx · 2026-10-15