Deploying on edge compute humbles you 😭
On the bright side, I have a pretty good grasp of the CUDA stack now.
Turns out an Orin Nano can actually handle flow matching policies
@Shadmanislam151@pentestduck@ycombinator You can arrange the arms/cameras like the YAM corpus they released and it transfers very well. Some people have FTed pi0.5 on it
@bennytay07@Scobleizer The FDE cost is too high for a lot of deployment companies. They all need insane $$ to build deployment teams. This results in only big players like Skild/Dyna etc actually deploying.
I really want to build the ecosystem towards being 3D printer level of self serviceability.
Forgot to take videos but demoed this today
Turns out deformable + hidden fragile rigid object (Maltesers lol) is a pretty hard challenge
Only 1 bag crushed :)
I’ve been getting quotes for humanoid parts lately. It’s made me realise that there’s a race to the bottom on arms specifically.
There are lots of options for decent QDD arms now <2k USD.
Horrified to find out even an entry level wheeled base costs >5k
PSA about the @seeedstudio B601
The stock leader arm (Star Arm 102) is quite hard to use and kinematically different to the follower.
Branching the U-ARM for the B601 right now - will see how it turns out.
GELLO also seems like a good option (the YAM config specifically)
Introducing GEN-1.5, a one-shot learner.
It can learn new tasks in a few seconds. Show it what to do, and it generalizes.
This capability emerged from pretraining on physical data at scale, as a step towards our mission of building general intelligence for the physical world.
Whenever I see an IL demo where they
1. Whip out a 🔦
+
2. Say it only needs an unbelievably small amount of demos
I get suspicious if it’s visual trajectory retrieval (vs trained policy). Nearest neighbour doesn’t care if the lighting changes 🤷
This is a non parametric policy with some ACT-esque tricks. Depth perception needs a little work (as seen by the gripper dragging). I’m only using a monocular wrist cam for this.
Seen some interesting demos recently from @ihorbeaver recently of imitation learning on extremely low demo counts.
Based on some hunches I put something together. This is a pick+place based on 9 demos. Can handle position + orientation change of 3-5cm ish.
Got remote VR teleop via Pi working :)
Like many people working with VLAs, I was tired of replugging 5 different things every session.
So I decided to abstract it all behind a Pi. Supports leader arm, Quest, DAgger
Disregard messy workbench - currently housing 3 projects 😭