Enabling Faster, More Capable Robots with Real-Time Motion Planning | Hacker News Reader

Enabling Faster, More Capable Robots with Real-Time Motion Planning

spectrum.ieee.org

41 pointsVarcht14 comments

They don't have any big breakthrough, just hardware-assisted collision detection. That's been done before, by PhysX, later acquired by NVidia.[1] The article doesn't say much about how their collision detection compares with what you can do with a CUDA graphics card.

[1] https://en.wikipedia.org/wiki/PhysX

We are faster than a CUDA graphics card. You get a win from designing custom hardware specifically for this task so that you don't have to load stuff on and off the GPU. The fastest previous approach takes seconds; ours takes milliseconds!

Stefanie Senior Director, Realtime Robotics

I have a robotics startup in the food space. Don't the majority of industrial robotics have a) ample space; and b) well defined motion envelopes? You mentioned suitability for "unstructured, dynamic environments"... so where exactly do you see this technology being used? I can see an argument for applications like high traffic area UAV ad-hoc flight planning (realistically that's mostly military or extreme specialist use cases such as firefighting in a burning building; in civilian use there's an argument for just systematizing the use of airspace in lieu of bothering), but not so much for production lines. I can understand the marketing challenge to explain an algorithmic efficiency advance to the general public, but getting a drink out of a fridge as an example just seems ... ridiculous ... IMHO humanoid isn't the future, humanoid is inefficient!

I can imagine a bunch of robots building a building and not colliding with each others work. You could do it without rt sensing but that's a lot of messages you could avoid.

Construction is all about operations management. You have to get it done ASAP because every day costs money in permits and equipment leasing, if materials or equipment aren't on site then the site goes idle, and storage space is often limited.

If you are going to involve robots, you would have them operate to a central plan, not invent their own. This means temporarily fencing off nominally exclusive spaces would be relatively straightforward.

Incidentally, I have an acquaintance who is heavily (tens of millions USD) invested in scaling a new business in the custom building domain. The concept is that a site is surveyed, a design set, then they pre-fab the building structure, ship it (internationally) flatpack in shipping containers, and - though it's manual for now - eventually assemble with robotics, an area in which they have been aggressively inventing and acquiring patents. I haven't asked but I very much doubt they are focusing on inter-agent motion planning as a priority.

It's complex enough, and that's essentially an avoidable problem.

Sure, they would of course need coordination just would not need to be told about every movement and addition. Think agile vs. waterfall.

Yes, but that is why it costs so much to deploy industrial robots. There are many applications where it is technically possible to automate today, but not economical, due to the cost of setting up the robot, due to it needing ample space and having to program a specific motion. If the robot could see and avoid obstacles and dynamically plan to pick stuff up, it would not need as much space, and it would be easier and faster to program them for the task (when the task changes).

Stefanie

Senior Director, Realtime Robotics

While adaptability as a device for expanding applications and market makes sense, your example of picking stuff up in a generic fashion also requires expensive and dexterous actuators, a reliable power source, plus knowledge of material properties and estimates of things like weight to facilitate an appropriate grip. Right now, that pushes costs higher than a manual deployment of dumb robotics (which will be generally faster, more reliable, and easier to maintain and hire for), which is why IMHO industry prefers dumb robotics. The reality is that most environments simply are not that dynamic, especially in industry.

Not only that, industrial robots are chosen for repeatability, not speed. A flexible system necessarily doesn't take the most efficient path, and if it does, you'd just buy a dumb arm.

Layman here, are there any serious limitations to your approach, or is it fair to say with the algorithms and hardware you have that motion planning has essentially been "solved"? Thanks.

developed special-purpose hardware to solve motion planning in under a millisecond

My poor dumb arduino bots...

https://spectrum.ieee.org/robotics/robotics-software/motionp...

This is literally just using an FPGA to accelerate the motion planning algorithm.

I like how the picture doesn't show the populated side of the board. Probably a high performance FPGA?

Lot's of talk about FPGA's here https://rtr.ai/technology/

There is a shot of the business side in the video I linked below, not sure if that is the same board.

Reply on news.ycombinator.com