New physics sim can compress training times from decades into hours

https://arstechnica.com/information-technology/2024/12/new-physics-sim-trains-robots-430000-times-faster-than-reality/

by elsewhen • 2 years ago

122 34 2 years ago

exabrial• 2 years ago

I believe this is how some robot vacuum companies perfected their algorithms 10-15 years ago; basically simulating inputs to the software from LIDAR/bump sensors and figuring out how to avoid getting the robot "stuck" in your living room.

ad-astra• 2 years ago

What companies are they? My Roomba from 3 years ago is awful

rogerrogerr2• 2 years ago

Curious, what makes you say it’s awful?

I have a couple roombas from that era. If I sit and watch them, their path planning makes no sense. But if I just put them on a schedule to clean once a day, and don’t think about them beyond emptying their bin, I have continuously clean floors. Which, for me, is all I care about.

pineaux• 2 years ago

Not enough comments, is this the ushering of a new era of robotics? Is everybody going to be replaced now?

amelius• 2 years ago

This piece was less about robotics and more about creating cool scenery.

kmoser• 2 years ago

Also, I thought it was pretty simple to write software to create random, procedurally generated worlds.

gnabgib• 2 years ago

Discussion (221 points, 3 days ago, 52 comments) https://news.ycombinator.com/item?id=42457213

aussieguy1234• 2 years ago

There are self driving car companies doing the same kind of training. I know because I spoke to an employee who works at a startup that provides this type of software as a service.

thatguysaguy• 2 years ago

Some cold water on the claims here: https://open.substack.com/pub/stoneztao/p/the-new-hyped-gene...

tl;dr: The way they're able to claim a massive speedup is by simulating much simpler scenarios than those being simulated by other systems.

elromulous• 2 years ago

Who would have thought simpler physics are easier to regress against /s

stonet2000• 2 years ago

blogpost author here.

This is not exactly the claim. I’m kind of certain they do “fair tests” where they test the same scenarios in different simulators. The problem however is that these tests are not indicative of anything useful. Why should one care if you get 430,000x real time speed simulating a robot arm not colliding or moving any objects with low sim accuracy? This number is fairly meaningless since if you do machine learning (specifically RL) to solve some task (manipulation, locomotion etc) that speed will drop significantly. Moreover in the past most simulators typically benchmark against previous versions of themselves and on scenarios with more collisions / somewhat more realistic situations (picking up a cube).

The original paper that started the trend of GPU parallelized sims is this one isaac gym: https://arxiv.org/pdf/2108.10470

They test on 3 environments, all with multiple collisions spanning locomotion and manipulation and get some nice speeds. These tests are far more realistic and grounded in reality than the test that made the 430K number.

amatecha• 2 years ago

Kinda nitpicking, but why is "SIM" allcaps here but not on the source article? It's a "sim" (short for "simulator"), not a SIM card for a cellular phone :'D

bongodongobob• 2 years ago

Probably posted on a MAC computer.

readyplayernull• 2 years ago

AI for the REST of US

solarkraft• 2 years ago

Threw me off too. Maybe autocorrect, similar to “APP”.

codetrotter• 2 years ago

Could be one of the HN auto corrects. It changes some words like that.

dang• 2 years ago

It shall SIM no more.

whilenot-dev• 2 years ago

OT, but that's the second time this month I'm encountering TLS issues with arstechnica.com... seems to be some mixup with the domains for intuit.com.

pipe2devnull• 2 years ago

When you read the fine print you find out that’s for a 2D simulation of a robot arm that isn’t moving. Real world results from coworkers showed it being about 10x slower than Drake in meaningful sims.

vagab0nd• 2 years ago

Sim is how we get to ASI. But physics is hard. In the meantime is anyone hiring a bunch of people to wear motion cap suit, cameras and mics 24x7?

New physics sim can compress training times from decades into hours

Related Stories