# State of the art telerobotics systems will not develop into general purpose telerobotics [[Working general purpose telerobotics is a system that enables a person to seamlessly interact with the world using a machine]]. It’s not as if people haven’t done a lot of work towards this goal, so it begs the question: “won’t we get there eventually?” I would argue that without intervention, the answer is either “no” or “far more slowly than we would get there with an intervention.” The burden of proof is then on me to argue why sweet demos like Jeff Bezos supervillianing it up, the [[Toyota T-HR3]], or the thousands of publications that come up if you type[“telerobotics” into semantic scholar](https://www.semanticscholar.org/search?q=telerobotics&sort=relevance).  [Caption] We dud ut! (Not really) In a nutshell, individual components have been optimized around a point in design space that would force the rest of the system to warp around them beyond the point of ‘working’ or were never created to be integrated into a system in the first place; telerobotic systems have either been designed for a hard-to-generalize use case, are a poorly-performing general-purpose robot with telerobotic components bolted on, or are optimized for demos with no serious context of use. ### Individual Components Most of the <commercial/built for a serious context of use> individual components that gesture at telerobotic possibilities are optimized around a point in design space that require a full telerobotic system to be built around their particularities. Adjusting them to play well with a system would require going back down the (abstract optimization landscape) hill and perhaps partway up another. Most robotic arms, like the ubiquitous Kuka arms are built for open-loop precision which makes them hard to use in situations where you might be bumping into things or handling poorly characterized delicate objects. Other arms trade off precision for compliance but without the ability to close the loop well with an operator, so they’re just floppy and vaguely bad. All the actually-used end effectors[^1] are some combination of pincers and suction cups meant to pick and place objects or specialized tools. Outside of commercial parts, there are many components that extrapolate quickly to general purpose telerobotics in your imagination but were never meant to be part of a bigger system. Instead, these components — from operator interfaces to new kinds of compliant arms to dextrous control demos — were intended to proof-of-concept a new way you could build that component or hit some benchmark for the component. The intention is that “someone else” will build the work into a larger system, but who? There’s also a class of components in a weird liminal space like the [[Shadow Robot Company]] hand or [syntouch tactile sensors](https://syntouchinc.com/) that I would call ‘research tools.’ They’re a product you can buy, yes. However, their main customers are people doing wildly varied work without a serious context of use. As a result, on the surface they do the thing (a dextrous hand! A tactile sensor!) but they just vaguely suck at everything. Additionally, while it’s extremely hand-wavy, the organizational DNA of the companies that make them just feel stagnant -- they’ve spent half a decade or more incrementally improving the same architecture and getting by on contracts. They have no slack to actually try new paradigms. ### Systems Current commercial telerobotics systems have been designed for a specific market that just doesn’t need general purpose telerobotics any time soon to be a good business. In these cases there are often architecture decisions that (despite narratives) would make it take just as much work to generalize the system (or modularize it to use parts of it for general purpose telerobotics) as it would to build a new system. For example, Cobalt Robotics cares a lot about a single person being able to operate many robots whose main job is to observe, interacting with the world just enough to unblock observation (by pushing elevator buttons). This is exactly what building security needs, but the company’s core gameplay loop doesn’t require generalized telerobotics so unless their margins go bananas, they’ll follow the incentives to continue optimizing the design decisions they’ve made. Nuclear material handling (just need to do a specific set of tasks through a lead wall), surgical (tiny slow static precise movements), and bomb diffusing (similar to surgical) systems are similar. On the opposite end of the specialization spectrum, many demo telerobotic systems (especially in academic labs) are a fully general robotic system with telerobotic components bolted on. These systems are generally a PR2 or Baxter robot with their cameras feeding to a person and some kind of control UI hooked up to their actuators. These robots fix so many pieces of design space that I suspect it is impossible to create a ‘working’ telerobotic system around them. Another, more recent class of telerobotic system is one whose purpose is to let a human help out a mostly-autonomous robot. This use case is serious and demands generalized technology that could be relevant for telerobotics but the architecture of the system is different enough that you can’t just tweak a few things to turn a human-assist system into a general telerobotic system. Finally, the most misleading class of system might be those created for the purpose of sweet demos. These systems are optimized to do a flashy slice of things that look and feel like they will generalize but haven’t been built in a serious context of use and as a result have fairly baked-in design decisions that have never run into the critical corner cases that would necessitate a deep overhaul. It’s also a mindset thing: the goal of the system is primarily to be a demo system — even when there’s are grander aspirations, you make different core design decisions when building a demo system than when building a working system that happens to generate good demos. I’ve been there. Most of the space telerobotics lineage falls into this category. Toyota is notorious for this effect in other robots (like Asimo) so I suspect that [[Toyota T-HR3]] is no different. I don’t have enough information to make the call on [[Converge Robotics Group]] (from the Jeff Bezos picture) but I worry that it might fall into the same category — created primarily to sell consulting work and components. Maybe this is too cynical and it just takes some more tinkering to turn these demo systems into useful tools. [^1]: The fancy robotics word for “the thing that grabs stuff.” ### Related * [[State of the field in Telerobotics]] [Web URL for this note](http://notes.benjaminreinhardt.com/State+of+the+art+telerobotics+systems+will+not+develop+into+general+purpose+telerobotics) [Comment on this note](http://via.hypothes.is/http://notes.benjaminreinhardt.com/State+of+the+art+telerobotics+systems+will+not+develop+into+general+purpose+telerobotics)