Comment by AareyBaba | Hacker News Reader

Comment by AareyBaba | Hacker News Reader

Parent Full thread

Whenever I see these Boston Dynamics video's I have many unanswered questions about the underlying control systems.

This one for example has an organic birdlike head motion. Is that pre-programmed or a natural outcome of their control system dynamics. The MIT technologyreview page says "Each one is created with carefully pre-programmed movements and will take many, many takes to get right before it’s shared."

So is it a sequence of individual actions: align to pallet using registration cards -> move forward with reaching distance -> detect box position -> use box picking algorithm to pick box -> backup fixed precalculated distance -> turn 90 degrees -> move forward to destination etc. This would be a brittle solution that has to be reprogrammed for the next task.

Or is it a more general closed loop solution: Theres a pile of boxes there. Move it here. Use your sensors to plan a trajectory and sequence of actions to accomplish the goal.

That's what I want to know.

I don't know much about control, but apparently Emanuel Todorov knows about the basics of Boston Dynamics control in this video https://www.youtube.com/watch?v=7enj1FGoYwg around 14 minutes. The paper is this one https://repository.upenn.edu/ese_papers/686/. I don't understand most of it, but it may help if you have some knowledge. But from the few things I understood, not much is programmed.

Also, I'm watching some lectures for a MIT underactuated robotics class (https://www.youtube.com/watch?v=VeEqtTgDXFc). The professor says in one of the videos that the people who end up creating Boston Dynamics also took that course. From what I watched until now, nothing is explicitly programmed (by some vague definition of "explicitly programmed"), you just find some inputs that makes the dynamics of the model behave the way you want.

Almost certainly the birdlike movement is the outcome of the robot dynamics, and not some hardcoded rules. But things like avoiding hitting obstacles or estimating the distance, I think that needs to be explicitly programmed (and then used as input to the model such as "move the end effector to that position").

I have no doubt that organic motion was deliberately designed that way to impress their intended audience.

Yes, it makes sense to mimic nature sometimes. But when it comes to robotic manipulators, loaders, etc - there's absolutely no reason to copy biological systems limited by strength, speed, or joint mobility. They seem to be deliberately going for that uncanny valley effect.

Eventually, someone is going to admit it. Until then, this is hugely entertaining.

I think it may very well be a natural outcome of inverse kinematics and PID loops in a noisy (uneven floor, uneven loads, etc, etc) environment

After looking at the video more closely, I agree with you. It looks like slightly underdamped damped response in a PID feedback loop. Funny, how it ends up looking like an ostrich.

yes, natural movements are approximately energy-minimizing. The optimal control algorithms used by BD probably have energy minimization as their main objective.

IK + PID is not how these robots work. These robots use MPC and trajectory optimization, which should absorb minor disturbances just fine.

Could you elaborate? The floors appear to be perfectly level, and load weight isn't changing after it's been picked up - so where's all this excessive locomotion coming from?

I agree that the conditions in the video are very ideal as you mention.

Perturbations do exist and they add up in the PID feedback loop. For example you can see in the video that the way each box is handled varies and that the robot reorients itself to pick/drop boxes.

Consider a simpler 1 dimensional example: https://youtu.be/JpNAhKT7yY4?t=101 (warning, loud). You can observe the system adjusting to the hits by the person, but also when left alone it will have to adjust to accumulating externalities like gravity. Also note the natural/biological looking swaying motion that results.

Thank you for explaining this. Still, while this presents a very interesting scientific problem, the engineer in me wonders why these feedback loops are needed for an automaton navigating a 2D terrain and carrying solid objects (unless they plan on tossing/catching boxes), why two wheels with that giant swinging counterweight? What's next, a unicycle?

You know, sometimes you see a solution, and it's just so elegant... This isn't elegant, or brilliant. This is an attempt to solve self-imposed, artificial problems.

P.S. people who liked your pendulum video (I did), might also enjoy this drone pole acrobatics video:

https://www.youtube.com/watch?v=XxFZ-VStApo

> This one for example has an organic birdlike head motion. Is that pre-programmed or a natural outcome of their control system dynamics.

Similar problems are likely to elicit similar solutions.

That’s the point of the question. Is this the result of convergent evolution or are Boston Dynamics simply mimicking nature’s solution? Even worse, what if the bird-like head motion has no functional value for the posed problem?

imho the motion looks entirely like a simple counter-balancing solution to keep the robot upright and the box steady. It does look like a whole-body solution, not a per-joint PID. Others have mentioned they use MPC which makes a lot of sense here. I think it's just "convergent evolution" if you want to put it that way.

I mean, isn’t that kind of Boston Dynamics’ shtick? When have any of their impressive demos ended up being useful in the real world?

Such discouraging words. Like literally everything a human has ever made, it takes _time_ to get it production ready.

They’ve been around for 27 years. BigDog was 14 years ago. Maybe “production ready” is not really a part of their company culture?

just wait for the next war. BigDogs and SpotMinis everywhere

Reply on news.ycombinator.com