Yes the problem is very hard. Mainly because high DOF generalization is very difficult.
We have self driving cars because what are the control inputs? Pedal, brake, steering wheel. This already took many many years.
Now for a humanoid robot: An action space that is metaphorically Hilbert. (Physically, yes, obviously)
Also, IMO, LLM's can aid the development of robots, but do little beyond a planning, human control interface. Below that it's the domain of control and the solution will be the correct combination of classical, neural, and real time optimization based control.
All the bad-ass biped robots that actually look natural? It's PID controls wrapped with control barrier functions constraining the QPs that are being solved in real time.
But that's annoying to derive per-application. So we'll need neural methods which can be learned (while being constrained by a priori knowledge of dynamics). My hunch is that the Yann LeCunn type of jepa models will be how tasks can be learned.
> All the bad-ass biped robots that actually look natural? It's PID controls wrapped with control barrier functions constraining the QPs that are being solved in real time.
That's not entirely true. Locomotion is well addressed by RL in sim. It's true that there is still a PD layer, and the RL policy produces setpoints for it.
>Yes the problem is very hard. Mainly because high DOF generalization is very difficult.
>We have self driving cars because what are the control inputs? Pedal, brake, steering wheel. This already took many many years.
Its actually amazing to me that this hasn't been solved yet. Its really not that hard of a problem.
Modern robotics, including self driving, are famously all about end-to-end training. We are trying to replicate what humans do through muscle memory. But muscle memory is not what makes us good at operating in the physical world. The thing that matters the most is our ability to simulate the world around us in a compressed form into the future, which lets us predict how our inputs will affect the world.
A similar system in a self driving car should be able to drive perfectly without self inflicted accidents 100% of the time, especially with basic lidar to serve as an error correction mechanism to the camera 3d scene reconstruction.
I don’t think the complexity scales with every additional degree of freedom like you are painting here. I think it’s just a matter of getting the right training data in sufficient quantities for an LLM to output across all degrees of freedom simultaneously without it being some exponential leap.
> So we'll need neural methods which can be learned
Data is a problem. LLMs had the advantage of the whole internet to train on. Robots don’t have that corpus of information. And real time learning seems to be something that everyone in AI is studiously ignoring.
The hope is that RL in simulation can fill the gap.
Also there’s imitating humans, via a suitable mapping from the human sensor, control and configuration space to the robot’s. Some groups have gathered video and other data from humans doing tasks, for example with a VR headset.
We have self driving cars because what are the control inputs? Pedal, brake, steering wheel. This already took many many years.
Now for a humanoid robot: An action space that is metaphorically Hilbert. (Physically, yes, obviously)
Also, IMO, LLM's can aid the development of robots, but do little beyond a planning, human control interface. Below that it's the domain of control and the solution will be the correct combination of classical, neural, and real time optimization based control.
All the bad-ass biped robots that actually look natural? It's PID controls wrapped with control barrier functions constraining the QPs that are being solved in real time.
But that's annoying to derive per-application. So we'll need neural methods which can be learned (while being constrained by a priori knowledge of dynamics). My hunch is that the Yann LeCunn type of jepa models will be how tasks can be learned.