An industrial robot's safety certification covers one exact configuration: change the gripper, the load or the task and the assessment has to be redone. That's already a modeling problem with hand-written pendant programs. Once the policy is reconfigured live, the way you describe with your local Qwen model and JEV-style control, there's no fixed configuration left to certify: the thing you're modeling keeps becoming a different thing.
The same gap shows up as a number. In a study of Fortune 500 retail and logistics deployments, warehouse robots trained in simulation dropped from above 92% success to as low as 67% on live floors within two weeks. I'd read that as robopsychology's problem landing as a metric: the abstraction stops tracking the system almost as soon as it ships.
stefanie, the unimate bit is the giveaway. devol filed the patent in 1954 and gm had the first one moving die castings in 1961, and for sixty years after that the robot's theory of mind was just reading the pendant program. now the policy is a black box and we are back to inferring intent off behavior, the way you would a stranger's.
Inferring intent off behavior is Dan Dennett's Intentional stance and goes back to videos such as Heider and Simmel's 1944 shape animation (https://youtu.be/VTNmLt7QX8E). But we don't get to just do it passively, we can also do it actively (asking questions etc). In the end, inferring intent off behavior is all we have in almost all situations.
heider and simmel is the right cite. the 1944 study showed 114 viewers a two-and-a-half minute animation of triangles and a circle, and nearly everyone came back with a story, the big triangle usually cast as the villain. the interesting difference is your active probing bit. you can't interrogate the animation, but you can interrogate the robot. that's new.
I think robotics has been a lot more sophisticated than just the teaching pendant since the 1960s. While a lot of researchers point to policy models being a black box, I don't think that is correct. They are entities we can dissect and evaluate. We have tools that can probe what an action head is doing, or what common patterns of statistics are going on in a VLM backbone. I think we can understand neural networks fundamentally much better than we can currently understand a stranger.
david, the one that keeps me honest here is representation engineering, zou's 2023 paper. linear probes that can read whether the model is lying straight off the activations. i don't have a probe that reads the guy next to me on the subway. the gap is just that most deployed policies don't ship with the probes attached.
In one of Leslie’s talk at IROS, she mentioned that the language we have developed so far is a fantastic way to specify the “compositional seams” of the physical world. And it feels like as we get new types of systems like LLMs to try and understand (I maybe would argue LLMs are not necessarily more complex than other things we already deal with), it totally makes sense that the abstractions we will use to understand them will be coupled with the language we develop. And potentially our language creation rate and capabilities have also increased. Something like UMI, I’m sure would not have been as widely known so quickly without social media, etc.
An industrial robot's safety certification covers one exact configuration: change the gripper, the load or the task and the assessment has to be redone. That's already a modeling problem with hand-written pendant programs. Once the policy is reconfigured live, the way you describe with your local Qwen model and JEV-style control, there's no fixed configuration left to certify: the thing you're modeling keeps becoming a different thing.
The same gap shows up as a number. In a study of Fortune 500 retail and logistics deployments, warehouse robots trained in simulation dropped from above 92% success to as low as 67% on live floors within two weeks. I'd read that as robopsychology's problem landing as a metric: the abstraction stops tracking the system almost as soon as it ships.
stefanie, the unimate bit is the giveaway. devol filed the patent in 1954 and gm had the first one moving die castings in 1961, and for sixty years after that the robot's theory of mind was just reading the pendant program. now the policy is a black box and we are back to inferring intent off behavior, the way you would a stranger's.
Inferring intent off behavior is Dan Dennett's Intentional stance and goes back to videos such as Heider and Simmel's 1944 shape animation (https://youtu.be/VTNmLt7QX8E). But we don't get to just do it passively, we can also do it actively (asking questions etc). In the end, inferring intent off behavior is all we have in almost all situations.
heider and simmel is the right cite. the 1944 study showed 114 viewers a two-and-a-half minute animation of triangles and a circle, and nearly everyone came back with a story, the big triangle usually cast as the villain. the interesting difference is your active probing bit. you can't interrogate the animation, but you can interrogate the robot. that's new.
I think robotics has been a lot more sophisticated than just the teaching pendant since the 1960s. While a lot of researchers point to policy models being a black box, I don't think that is correct. They are entities we can dissect and evaluate. We have tools that can probe what an action head is doing, or what common patterns of statistics are going on in a VLM backbone. I think we can understand neural networks fundamentally much better than we can currently understand a stranger.
david, the one that keeps me honest here is representation engineering, zou's 2023 paper. linear probes that can read whether the model is lying straight off the activations. i don't have a probe that reads the guy next to me on the subway. the gap is just that most deployed policies don't ship with the probes attached.
In one of Leslie’s talk at IROS, she mentioned that the language we have developed so far is a fantastic way to specify the “compositional seams” of the physical world. And it feels like as we get new types of systems like LLMs to try and understand (I maybe would argue LLMs are not necessarily more complex than other things we already deal with), it totally makes sense that the abstractions we will use to understand them will be coupled with the language we develop. And potentially our language creation rate and capabilities have also increased. Something like UMI, I’m sure would not have been as widely known so quickly without social media, etc.