What's the actual bottleneck right now: hardware or software?

[kwilliams]

@servoken70 Just to be precise about one thing: Open-source humanoid projects matter less for out-competing the well-funded players and more for keeping the barrier to entry low enough that new ideas and new people can actually get into the field. A recurring pattern in the field: a capability looks solved in a controlled demo, then takes another 18-24 months of engineering to become boringly reliable in the field.

[zoeanderson]

@kwilliams Ran into exactly this myself. Legs earn their keep specifically where the environment wasn't redesigned for robots - stairs, curbs, uneven yards, cluttered floors - which is a smaller slice of real-world tasks than the marketing suggests.

[jonathan.rao1]

Just to be precise about one thing: A lot of the current wave genuinely is repackaged 2000s and 2010s research (ZMP walking, whole-body control, DARPA-era grasping) combined with two things that weren't available then: cheap high-torque-density actuators and large-scale learned policies. A genuinely underrated bottleneck is neither hardware nor pure ML - it's the unglamorous integration work of making perception, planning, and control agree with each other reliably for eight-hour shifts.

[kwilliams]

@jonathan.rao1 Respectfully, I think this undersells it a bit. Wheeled-torso designs solve 90% of warehouse and light-industrial tasks with far less complexity than a biped, which is exactly why so many 'humanoid' demos quietly have wheels under a skirt.

[servoken70]

@kwilliams I'd frame this differently. A genuinely underrated bottleneck is neither hardware nor pure ML - it's the unglamorous integration work of making perception, planning, and control agree with each other reliably for eight-hour shifts.

[pierregreen]

Slight correction, though the overall point stands: Legs earn their keep specifically where the environment wasn't redesigned for robots - stairs, curbs, uneven yards, cluttered floors - which is a smaller slice of real-world tasks than the marketing suggests. A recurring pattern in the field: a capability looks solved in a controlled demo, then takes another 18-24 months of engineering to become boringly reliable in the field.

[jhansen]

@pierregreen I'd push back on this a bit. Elder care is one of the most emotionally appealing use cases and also one of the hardest, because it demands both physical safety around fragile people and social/communication skill that's still far behind the manipulation hype. A genuinely underrated bottleneck is neither hardware nor pure ML - it's the unglamorous integration work of making perception, planning, and control agree with each other reliably for eight-hour shifts.