Hacker News Reader

Comment by greenchair | Hacker News Reader

question for the robot experts: what is the limitation that makes the movements so slow? for example when it picks up the ball and puts it in the basket. why couldn't that movement be done much faster?

View on HN

From university, I vaguely recall that, I had to implement a lot of feedback and correction calculations when working on industrial robotic arms. Usually too much speed causes overshooting(going the wrong trajectory or away from target). The feedback is constantly adjusted until the target is reached, hence a lot of expensive computation and readjustment from all the sensor feeds. Additionally, faster movement also has risk of damaging nearby objects when overshoot happens and also harms/degrades the joints faster. For a simpler example, think about the elevator, what would happen if it were to go up/down very fast, how would you tweak your PID controller to handle super fast movement to not throw your passengers when you need to correctly align and halt at the target floor….

Camera feed processing latency would be my guess. The system needs to make sense of a continuous video feed so moving slower reduces how much happens in between frames.

In this case it’s the model. There’s an insane amount of computation that should happen in milliseconds but given today’s hardware might run 10 times too slow. Mind you these models take in lots of sensor data and spit out trajectories in a tight feedback loop.

I’m no robotics expert, but look how close the robots are to squishy human meat bags.

I assume Google is being very careful to keep the speeds well below the “oops, it took your jaw off” threshold.

This is highly unlikely to be a mechanical limitation of the robotic arms. As others have said, it's likely an inference speed limitation - their model is understanding, reacting, and producing outputs as fast as its supporting hardware can.

But that all just poofs away in a year or two as inferencing hardware gets better/faster. And for many use cases, the slowness/awkwardness doesn't really matter as long as the job gets done.

"AI working in meatspace" was supposed to be hard, and its rapidly becoming clear that isn't going to be the case at all.

Robot demos have been stuck at that sort of speed for more than a decade and they didn't have to wait for a giant LLM to do inference. Why's that going to get any better now? And just in a year or two?

F=ma. An arm that's powerful enough to move extremely quickly is powerful enough to hurt.

I'm not a robot expert, but I do know the answer is simply safety. Once it learns what to do, it can do it faster and faster, but when something goes very wrong, it will go very wrong.

inference speed of the models is probably the bottleneck

Reply on news.ycombinator.com