Posted by xnx 4 hours ago
Based on my very limited knowledge... if you would have dexterous hands with sensitivity on them you could theoretically train a VLA model to do that task, planned by a high-level model like Argon 4, that orchestrates everything. That would as well require in first place a VLA that has similar tasks in its training set and fine-tune it by means of recording hundreds of examples performing teleoperation of the robot hand and arms movements.
Not easy, but definitely getting more and more doable these days.
PS: Didn't spend 20k on hands for my 6k robot, but can confirm it has the cable harness to support them :D
What's the latency like? Is the llm used to create an instruction set for the robot? How is it interrupted for reflex actions like controlling stumbling?
In any case latency is acceptable for navigation I guess, but avoiding moving obstacles is definitely a no-go right now.
It's 1.23 and weights ~30kg. Even though only the R1 Edu can be in theory programmed, it can be easily modded to enable SSH and root access to deploy custom movement policies and interact with the different on-board daemons to control the movements, grab the video data, etc.
I bought a second-hand Intel DeepSense camera and plugged it on one of the spare USB ports available in the main board to have 3D sensing (even though I really need to play a bit more with it). Right now I am playing with MuJoCo for RL training of new policies, integration with VLAs like NVIDIA Isaac GR00T and integration with LLMs.
I have to say its being a wild ride, and despite the small size of the robot, I have already injured myself with it when playing with the walking policies (If you activate a policy while its hanging, it starts literally kicking and punching you...) I cannot image those owning with the bigger brother of it, the G1...
- Zero shot control using a visual harness and action primitives. I've seen this from a few other places as well (https://tml.stanford.edu/homebody/) and it's essentially having a generalist LLM command a robot via a set of predefined actions. Many robots and many tasks can probably be done with this approach, since not everything needs super intricate 6DoF arms or dexterity
- simulation bring-ups: I have a pipeline that takes in satellite imagery and reconstructs the lots and houses in a 3D environment and snows on it so that my virtual robot can learn to clean. With these latest models, you can probably create a higher fidelity digital twin more directly.
- self-improvement: These models (Argon, Astra, etc) can't operate the robot itself, but would be able to create data to distill into smaller models you can actually utilize.
Aren't we still like a decade away from an Argon model being able to run on battery power for any reasonable amount of time that would fit on a robot that isn't the size of a school bus?
Add on another couple of years for the speed to get to what it likely needs to be to be generally useful instead of narrowly useful?
I mean - I guess we're maybe 5-7 years away from a reasonably priced robot you could plug into the wall like an old vacuum cleaner (that runs on this type of tech).
My assumption is the hardware will have one/several smaller models running on it for locomotion, balance, etc.
Then the hardware "phones home" for decision making and planning.
Essentially the planning doesn't need to be close to the device as long as the local models can handle some part of the plan execution and make sure the robot doesn't fall over. The latency shouldn't be an issue. In fact more powerful centralized compute quickly creates an advantage over slower local compute in terms of end to end planning.
Sooner or later, the models will stop improving. After that happens, the best models will be etched into model-specific ASICs with weights fixed in mask ROM (like Taalas is already doing), which will result in massive, at least order of magnitude inference efficiency increase.
And although the industry hit a wall in density/area, they switched to scaling in 3D or glueing chiplets together. Moore's Law is still working if you change the definition to transistors/integrated circuit.
Robots stand-alone AI do not really benefit from things like batch mode so you want to keep your LLM as high level as possible and your robot's subsystems as autonomous as possible, very light coupling between the two, definitely nothing time-critical. And that's the challenge: in robotics almost everything is time critical.
Also, your model probably doesn't need like a lot of its knowledge to control a robot.
From a cost perspective, I just can't imagine that heavily, heavily specialized models wouldn't crush a general model, though - which Amazon already does.
Robotics tends to be at the worst possible intersection point of all three of those: limited power budget, relatively small model size and cheap enough to be affordable when embedded into an appliance.
It just seems like jumping the shark a little bit.
It's basically completely unviable for a few years still.
I don't think it's viable in the sense that they are still in a demo stage.
Also, I would assume these models are going to get bigger before they get smaller (but they will get smaller over the long run)
For a set of tasks that almost no one would use a robot for...
I think someone should make a game based on this concept.