Note that the main control loops are running at 100Hz, and the adaptation is running at 10Hz. That makes sense.
I was working on this problem back in the early 1990s, when we had to do it via control theory. This new way is better. Having a few more orders of magnitude worth of compute power, (the robot seems to have a Raspberry Pi 4 on board) and 25 years of machine learning progress helps.
It seems to be purely reactive, though. They haven't yet moved to running with significant air time, which requires projecting ahead to the landing. That's been done for some special cases, mostly backflips and such, but the agility of a running horse or a basketball player will require projecting ahead.
This will come. Enough people are working on it now.
Classical control theory is often sufficient for dealing with changing loads, terrain and wear and tear. Rejecting unknown disturbances is what a control system does. No training is required. Perhaps somewhat less than impressive due to the choice of robotics platform, but:
Yeah, a lot of this tends to be "Let's try CNN on ... ", and it ends up doing about as well. It's bound to be a general purpose tool that's easy to use and doesn't require a classical approach.
Like RRTs instead of visibility graphs for regular-old polygonal environments. The RRT is just easier vs hand-tuning obstacle boundaries, etc.
I have some experience on this, as we are building a system like this with similar constraints, and there's really no "need" for anything other than classical control.
I went to a lecture by a professor at Oregon State University where he discussed research into bipedal walking, and described much the same problem and how you solve it. Namely, how do you adjust to changing terrain and payloads.
One of the take-aways from that lecture (and this is my naïveté showing) was that their research (that later produced the “Cassie” robot and the company Agility Robotics) was you can either mimic nature and how animals walk, or you can do it like Boston Dynamics and “Brute Force it”
Am I the only one who busted out laughing, repeatedly, at the totally comical failure modes for the original manufacturer's firmware? At t=1:19 onward. That is one hilarious butt-in-the-air faceplant at t=2:17.
Not only does it fall over/forward/wherever, it manages to look sad and helpless in the process :) Especially liked the first one where, after falling off the course, the OEM firmware just sort of dejectedly sinks slowly to its knees and lies down on its side.
> the OEM firmware just sort of dejectedly sinks slowly to its knees
You're almost certainly seeing the first moves of a preset 'stand up after falling' procedure that goes into a 'loaf' position to roll into the right orientation and get all four feet in contact with the ground.
I love how animalistic it looks - you could readily give it a realistic dog costume and convince people you had a live pooch (maybe a bit weird, but breeding has produced far less grace in motion than this thing.) I'd like to see it going up and down regular stairs, though.
The A1 bot they're using is $10k, less than some purebred show dogs - maybe we'll see consumer wallet friendly versions of these in the next decade. Rote yardwork, cleaning trash and sidewalks, clearing animals and bugs from crawlspaces and tight areas - heck, a "pack"of these could probably help animal control folks herd strays.
Even if battery life is only 30 minutes between recharging there might be a valuable niche for these guys. Imagine festival and event cleanups, retrieving needles and glass and other unsavory or unsafe hazards would be much safer and more thorough.
These could have a rich platform for tons of applications once the navigation is solved, and this looks like a big milestone.
How creepy would it be if they mounted a toy animal head to the front of the robot?
It strikes me that "balance" is probably one of the rules of legged animals (thinking of rules in terms of the boids rules of separation, alignment, and and cohesion).
Fascinating to think of the unfolding of complex systems from DNA to brains and nervous systems to instincts - programs decompressing over many iterations into programs that contain all the functions living things use in the world. And then, looking at the deceptively simple and repetitive structure of the brain, to think that evolution came up with a single program with infinite depth and sophistication to solve whatever gets plugged in.
And then it takes intense dedication and education to replicate one behavior, like walking.
I'm not sure if it'd be creepy. It looks to be pretty solidly out of uncanny valley territory... I think I'd believe it was biological. These guys did a phenomenal job.
I'm always struck by how it looks like a blindfolded animal blundering along. I hope soon we get forward-looking vision systems that can classify hazards and terrain types to allow thinks like leaping over mud and accurate footfalls. That's what living things really seem to do.
Note that the main control loops are running at 100Hz, and the adaptation is running at 10Hz. That makes sense.
I was working on this problem back in the early 1990s, when we had to do it via control theory. This new way is better. Having a few more orders of magnitude worth of compute power, (the robot seems to have a Raspberry Pi 4 on board) and 25 years of machine learning progress helps.
It seems to be purely reactive, though. They haven't yet moved to running with significant air time, which requires projecting ahead to the landing. That's been done for some special cases, mostly backflips and such, but the agility of a running horse or a basketball player will require projecting ahead.
This will come. Enough people are working on it now.