Do you think there is a path/we are making progress toward less leaky robot abstractions? I feel like I could point to actuators being potentially less leaky with their commoditization (maybe also manipulators), but then higher up the stack I don’t think we have centered around sufficiently non-leaky abstractions (maybe ROS? Maybe simulation toolboxes? RL policies?) is robotics doomed to be forever leaky 😞.
On a side note, there have been a number of attempts to circumvent the whole explicit abstraction challenge by just having a single entity deal with everything above a certain level of abstraction such as VLA, Astra, etc. is robotics so inherently leaky that the only way to deal with the abstractions is to not deal with them?
Krishna, the uncomfortable version is that the leak just moved up a layer. Brooks was already refusing the abstraction in 1990 with Elephants Don't Play Chess, wiring sensors straight to actuators with subsumption. Thirty-six years later the same leak reappears between the VLA and the physics, except now you discover it at 2am instead of in design review.
I think AI is a benefit and a curse to the law of leaky abstractions. The benefit is that it can relentlessly probe the software part of the codebase to find and fix bugs, especially ones that are software verifiable like math is. The curse is that it can stand up a much larger software stack much more quickly which itself creates more bugs. But as we are finding in my robotics class this semester, it can't help (yet) when you soldered a short or when a propeller cuts a wire.
the soldered short is the giveaway. pichai said back in 2024 that more than a quarter of new code at google was ai generated, so the machine writing the stack is also the one auditing it. fine until someone solders the short, which the model can stare at forever and never see.
Do you think there is a path/we are making progress toward less leaky robot abstractions? I feel like I could point to actuators being potentially less leaky with their commoditization (maybe also manipulators), but then higher up the stack I don’t think we have centered around sufficiently non-leaky abstractions (maybe ROS? Maybe simulation toolboxes? RL policies?) is robotics doomed to be forever leaky 😞.
On a side note, there have been a number of attempts to circumvent the whole explicit abstraction challenge by just having a single entity deal with everything above a certain level of abstraction such as VLA, Astra, etc. is robotics so inherently leaky that the only way to deal with the abstractions is to not deal with them?
Krishna, the uncomfortable version is that the leak just moved up a layer. Brooks was already refusing the abstraction in 1990 with Elephants Don't Play Chess, wiring sensors straight to actuators with subsumption. Thirty-six years later the same leak reappears between the VLA and the physics, except now you discover it at 2am instead of in design review.
I think AI is a benefit and a curse to the law of leaky abstractions. The benefit is that it can relentlessly probe the software part of the codebase to find and fix bugs, especially ones that are software verifiable like math is. The curse is that it can stand up a much larger software stack much more quickly which itself creates more bugs. But as we are finding in my robotics class this semester, it can't help (yet) when you soldered a short or when a propeller cuts a wire.
the soldered short is the giveaway. pichai said back in 2024 that more than a quarter of new code at google was ai generated, so the machine writing the stack is also the one auditing it. fine until someone solders the short, which the model can stare at forever and never see.