Getting a Leg Up

Please give a warm welcome to CEO Damien Shelton CTO Melanie wise and our special Guest digit [Music] Hello thank you everybody for coming so Robots with legs have seem to have had a Really slow start but suddenly they're Absolutely everywhere now humans are Pretty neat we have all these arms and Legs and stuff but planes don't have Legs planes cars don't have legs why do Robots need legs I think the the short answer is because Our environment is constructed around us So if you think about requirements Documents you can kind of run that both Forwards and backwards and if you start For an environment that was designed for People a robot that can move easily Through that environment would end up With arms and legs so the stage was Built for humans shall we see if Digis Can join us out on stage here sure Awesome so while uh digit is strolling Out on stage Um I noticed so I'm I met one of the Robots the other day and I notice it Seems like it has eyes It doesn't have eyes yeah that was a Controversial decision actually with the Original digit V3 that we started Selling three years ago some people saw It as not having a head I was in the it Had a head camp but then we were looking

At the human factors side of it one of The nice things with having the eyes is It gives you a sense of where the Robot's about to go so if you actually Watch it when it turns it's going to Kind of look in the direction uh of Where it's going it can blink it can do Other gestures and that gives you a nice Human factors interface to the robot Right but it doesn't actually use those Things to look right the sensors all all Everywhere so it's a Communications tool Correct yeah What what like how how does that show up In terms of uh how it kind of moves Around and and greets people uh and uh Interact with the people around it yeah I think that when you look at digit the The human robot interaction concerns That we care about or around making People feel comfortable and uh I guess Connected with the intention of the Robot as much as possible yeah so that's Why digit has eyes obviously Um but also it helps to direct the human Gaze to areas of the robot that are Important for them to care about for Example the LED ring that's on the sides Of the head and things like that so that People see digits in a good State and Excited to be in the world right And I think that as as digit does more Things with our customers in the world These capabilities were will evolve yeah

To help communicate different things to The customers So what took me a little bit through Through the gestures and things it's Doing right now like what what do these Uh represent to you as a roboticist yeah I think that when you look at digit Digit is really built for work and it's Built for work in logistics and Manufacturing facilities and so a lot of The different capabilities that digit Has are to accommodate those Environments so being able to reach very High being able to crouch down very low Being able to move its arms in a way That allows it to reach around big Things Um and all of these things are super Important you know as someone who spent 10 years in the AMR industry building Mobile robots without legs right Um there's there's quite a few Limitations that you you start to run Into with statically stable robots and With digit being dynamically stable We're able to reach a lot more of the The workspace of the facility that People can reach right and so you talk About Um spaces that are built for humans so That this robot is good at interacting With and in a lot of spaces obviously They're kind of mostly built for robots Right it's for robots around so is this

A stepping stone towards something else Or do you think that these types of Robots are needed throughout the kind of Yeah I mean I think you know a question We got early on is do you imagine digit Grabbing something and then running down An aisle carrying it and I think in most Practical deployment scenarios you're Actually looking at sort of a Heterogeneous robot deployment so digit Would work alongside an AMR which is of Course Melanie's whole background I Would work with other static fixed pick And place robots right so the goal is to Infill automation right now in ways that Can only be done with people who have a High injury rate this is a an industry That has historically uh very high Underemployment in fact growing Underemployment as uh the push for Logistics increases so very much this is A part of us Universe of solutions but Not by itself the complete solution all Right show away goodbye to uh digit Folks five digit Thought so goodbye Um so what was the hardest part about uh What we see in front of us here like What was the the biggest challenge of Developing a robot like this uh what's The hardest part what's the hardest part For me are the poor Engineers who Actually have to build the road yes uh You know I think what was interesting

You know our first time at TechCrunch Was at the robotic sessions I think five Years ago I was there and that was yeah That was with the Cassie robot and so I Think what we had to do was plan out a Structured approach of starting with What was unapologetically not a Commercial product right we sold it to Academic r d and we had to kind of Figure out a way to ladder ourselves up All the way to now a generation of Robots that's deploying into customer Environments for me that whole long Arc Was the hardest piece but yeah and we're Now just starting the ramp to the really Really really hard part which is Productization right uh it's one thing To do something well once it's another Thing to do it well at 99.98 of the time time right and that's A a strong and long ramp and we're Excited to be on it and that's that's The biggest thing and that's why we had The announcement of the robofab yeah um This week and so for context I wrote About it but you were there so what is a Robo Fab yeah so that's our new facility That we're opening in Salem Oregon which Is north of our current engineering Facility of about 30 miles it was a Building that we happened to intercept At the right time in its construction so We started this project about a year ago And the reason if you've ever done Home

Remodeling it rarely finishes in a year In this case we happen to get in with The landlord at the right time to bring The facility online right now we're in The process that I you know a lot of uh EV makers go through early in their life Where you're going from initial Mid-volume hand-built robots up to a Transition to a line approach which is What we're doing uh later this fall and Through the spring with the beta version Of Digit version four yeah our digits Being used to build digits not yet but We are are going to be using them to Carry totes around in our facility so we Do kidding when we hand off to the tax Uh the parts list to build the robot That goes into a bin turns out one thing Digits very good at is building or Moving uh bins around facilities yeah What's the like what is the Um how much can it lift how how can it Lift things yeah so if you if you look At digit today Um digit can lift about 35 pounds uh it Predominantly moves totes I can get up To 1.4 meters and it can get down to Basically the ground Um in terms of its ability to reach and Grab things it seemed like it was Reaching much higher than 1.4 meters Just now yes it was but I think the the Thing is is that when you look at even 1.4 meters or 2 meters like OSHA has

Some limitations on how much weight a Person can lift and so even though digit Can do those really high lifts we most Likely wouldn't encounter those lifts in A facility that was built with people in Mind right that makes sense that makes Sense so um Like backstage there was somebody making A joke about Skynet and I'm sure you get Those seven thousand times a day Definitely not that will happen yeah it Never happens I thought so uh but Weaponization of robots and stuff is a Conversation that people have a lot and I know Um you guys signed a A pledge sure yeah uh what was that Pledge and why yeah so we we were we Actually know the folks at Boston Dynamics really well it's a relatively Small community uh and uh we had started Talking with them about a year and a Half uh ago and started talking through What do we want to see the industry Evolve into uh and you know something That all of us are quite sensitive to is The desire to not have a robot ever hurt Someone whether accidentally or Intentionally the easy thing to prune Off is like no intentional hurting of People there's just no reason to do that It doesn't help the industry Okay Life it's very hard to see a Universe

Where I intentionally hurting people is On that road map yeah What does that actually specifically Mean does it mean not attaching a weapon To a robot like what is the Pledge what Are you pledging so Something Not using the robot in any way that can Harm people or animals or even be Threatening yeah and so what why is that Important to you or like what happens if Your board comes to you and says hey Actually Thank you Like uh I'm in Tech because I think it Can do good things for people and if I Didn't think that was the case I Wouldn't do it um you know the I think At the end of the day you just have to Decide why are you in the business uh You know I people can make decisions on Their own but for me personally it's Just not something that I would be Comfortable doing the business I uh I've done a lot of different things I'm not uh someone who had a life's work Uh Melanie has been a robotics Certainly has been in robotics uh my Goal you know my wife and I always sit Down before I do a startup this is my uh Second uh and the first time she said What's your goal I said learn a lot Don't go broke uh which is always a Worthy goal I'm sure as a typical

Founder's advice guys Uh for this though it was go big or go Home and specifically I wanted to see a Fella thousands and something that no One ever wow 10 000 a year that's the peak capacity yeah Right that's more than a thousand yeah It seems like that yeah yeah Um why is that exciting for you to build A scale Yeah yeah I mean um Before you know in Fetch uh Robotics and Um you know I've already been through The Journey and I really loved it and I Wanted to do it again Um and it's I've already decided a Journey with kind of the the next Evolution of kind of Robotics automation For Logistics so you know robotics Automation has been going on since the 1960s so we're about 60 years in and we Went from fixed Industrial Automation to Mobile which I spent a lot of time in And now we're doing mobile manipulation And I I think it is the future and I'm Super excited to be part of it yeah in a Lot of the kind of broader automation Industry there's a lot of conversation About jobs job security scaring us Around that you know um I think most Recently I saw that six percent of all People in the US work in trucking well a Lot of those jobs are about to go Bye-bye

Um and that sucks for those people so There's retraining there's early Retirement of all those kinds of things This robot starts to look an awful lot Like a human being Um what's your take on that like what Yeah so Um so I spent the better part of the Last decade deploying mobile Robots right in logistics and when I Started in 2014 with fetch robotics uh There was a 600 000 job Gap so that Means that they could not find 600 000 People to do the jobs and Logistics and Manufacturing in that time the Gap has Grown to a million people so and in all Of the time that we deployed robots at Fetch robotics we never replace the Person because every single time we were Just trying to fill a Gap a growing Never-ending Gap and so I I don't see This like this problem happening right Now because the workforce is aging People don't want to do these jobs and The numbers are showing that annual Growth in you know the job Gap is Getting larger and larger in the United States and abroad yeah and so when you Say these jobs what what are the jobs That that digit is able to fill that That currently people aren't uh yeah so The Bureau of Labor Statistics the BLS Report includes material movers Freight Movers hand manual movers so these are

People that are are moving pallets of Goods moving individual Goods uh doing Picking activities basically anything That's required to get a an item out of A facility like at Amazon or or any of These places from a truck to a shelf to A shelf into a box onto a truck into Your house so that that's what's Encompassed in all of the BLS data for Material movers and hand workers so You're mentioning Amazon and I imagine That if you are Amazon and you're Building you're building a brand new Warehouse you're probably going to build It mostly for the robots so where does Digit fits in for that or maybe I'm Wrong feel free to correct me uh go Ahead yeah I mean I would say the the Interesting answer is no uh and you know Certainly Amazon's an investor of ours So just be clear I'm not speaking on Behalf of them uh a lot of these Facilities right now lack good ways to Glue automation from different system Providers together uh and a great Example of this was the demo that we Showed at promat which was pulled from a Customer application where we had a Gravity uh feed of toads that we had to Move to a conveyor you'll end up in a Lot of these cases with very specialized Very highly optimized solutions for These material moves that aren't by Themselves attached to each other and so

When you set up a factory interestingly Even in a Greenfield construction site You're going to end up with these gaps That right now are filled by hand labor Yeah yeah and digits kind of the Automation glue between the automate Islands of automation when you really Look at it it's really helping connect Those disparate pieces like Damien said Yeah so I didn't panel earlier today on This very stage about uh intrinsic and Their work around connecting like I feel What they're trying to do to build a Kind of Hardware agnostic approach to All of this and whenever I say Hardware Agnostic That's the response I get from robot Assists Um what do you need to do to program a Digit Do you want should I take this you are The CTO yeah Well Um so right now Um since we're in early deployments with Customers there is a lot of black box to The to the deployment we're working Towards uh more visual programming Languages like intrinsic Um that help basically distill down the Workflow logic of go here do this into Something that is easily composable and A lot of other companies have done a lot Of work in this space and so this isn't

Something that is you know Groundbreaking rocket science Um I think that the big thing is is it Comes down to long-term customer Engagement and getting to the parts that Are the most meaningful to the customer Because digit as you saw has a ton of Capabilities and making sure that we can Uh get to the right level of Composability for the workflow that the Customer cares about is what we're kind Of dialing into okay yeah I can tell you About a Bad Bet that I made one time I Would love to hear about your bad bet uh The Bad Bet was that we could think of The robot like a smartphone where people Were actually pushing binaries to it and So we had originally structured some of What we called the applications Engineering team with this eye towards Composing monolithic applications with The eventual thought that well maybe This is exactly what customers would do Over time think of how smartphones have Evolved right so people forget that the IPhone didn't launch with an app store But now it's sort of the core feature of Smartphones we've changed our thinking And when Melanie's talking about The Primitives we now actually don't even Use apps as the parlance when we're Talking about it we call it skills now Where there's these much much larger Building blocks that abstract the way a

Lot of the core requirement of quote Needing to code on the robot right and Make it a whole lot more generalizable Right yeah I mean I don't want to Control each individual actuator on a Robot it wouldn't be able to stand up Right at some level I have to be able to Tell it like like can I program it uh Yeah you can actually so uh Bambi one of Our senior software Engineers who does Human robot interaction actually had Some Elementary School students Programming the robot using a visual Programming language I think uh they're Probably smarter than me so it's not but That sounds really cool yeah it was what Level what level of programming is it Like go over there pick up bucks put box There or exactly it's all those skills Primitives yep Cool and to what degree can it kind of Do stuff autonomously it's like hey go Pick up speaker put speaker over there Kind of thing look or do you have to be Pretty specific about what the task is So Digi is a an autonomous agent right And so what we end up doing is part of The workflow construction is we give it Like a task list right and we say these Are the tasks that I need you to do and They're parameterized so typically the Way this works in industry is like There's the WMS this cloud-based or Whatever system that holds all the

Magical data that tells you where Everything is and so the wwms number uh Warehouse management system or you can Have a warehouse control system they Have all sorts of acronyms the alphabet Suit is large and and vast Um but when you get down into it that System the system this black box system Talks to our system and we basically Have a workflow that says go here do This and basically the go here the here Part is populated by communication from This other system right and so when you Have a workflow that is digit goes to a Shelf takes a toad off of it and brings It to a conveyor basically the WMS is Saying this is the tote and then when Digit does that then it says oh this is The new tote and then digit does that And so you can you can imagine a set of Instructions that go to digit which is Get tote where tote is from here Walk to conveyor place on conveyor get Empty tote put back where you just took That tote grab new tote start again and So and that's a very simple way of doing It but it allows you to do some very Complex behaviors with the robot yeah I Feel like one of my first jobs was Actually in a paint factory and I was Instructed exactly like that get Barrel Move barrels like yeah that's cool and If and this is the thing the person who Gave me those instructions would not in

Theory have been able to program a robot But if it or program you personally like They're not going to Puppeteer you right Right well the robot is not going to sit On the stage 20 years later and tell a Grumpy story about a bad manager Hopefully but there is a there is a Piece to this where even the people who Don't necessarily know how to program Robots are able to program robots and I Kind of like and the more the robots Start to look like a human like it's Roughly my height roughly my weight Roughly my everything like that becomes Really interesting because you you end Up having a very similar relationship With these robots that you would have to A to a worker perhaps yeah we see you Know we get as I'm sure everybody does 20 questions today about are you using Llms for anything right and are you Using lens for anything uh in the full Deployments no uh because AI Observability right now and safety Issues uh mean that we're not using Those techniques outside of a few very Limited cases so there are some machine Learning RL approaches that we use on The robot and in-fold deployments but Not uh generative AI but the interesting Thing about this and there's a video on Our on our YouTube channel that we Showed some very early rough work with This is precisely that that if you layer

Uh a modern generative AI on the top end It's not puppeteering the robot the data Rates are way too low to do that out of Even relatively good llms these days but What it can do is provide that task Specification get Barrel carry Barrel Here and and compose more natural Language tasking into a set of commands The robot can execute and so how far are You away from like is that something That is desirable even in the long term Probably right uh in the short term Again for safety and observability Reasons now And when you say observability me does It mean that if the robot does something It's not meant to you don't know why Correct That's a classic AI problem Right and so how does that show up like What are what are you worried about Well I mean I I we're probably worried About different things but uh in my case You know we do have standards you know We're we are right now the world leader In in robotics and have been for the Last three years and at least in so far As humanoids are concerned so I think You know like with the weaponization we Had to have the opportunity to decide What the future looks like and one of The things that we want to do is say no We're not going to be irresponsible About this we want to roll the robots Out work with our customers and the

Standards agencies to get robots safely Into environment one of the Characteristics of that kind of Successful action is being able to tell People here is exactly how safe the Robot is so in some cases that means we Put a laser fence around it if people Have access to it in other cases it may Be very safe to be next to the robot It's going to be very task dependent and Very specific and we don't want to leave A lot of that open to you know AI Systems that even you know people who Are much smarter about it than I do kind Of head scratch say well I'm not Entirely sure why I drew a picture of a Cat when you asked for a mouse right um So yeah that's my answer well I'm glad You say It's relatively safe because we Just squeezed by I was like What scares you yeah so when you when You Advocate Learning Systems the the Biggest thing is is can you for the Things that matter like safety can you Have complete determinism because you Want to know that it's going to do the Right thing every time and so one of the Problems sometimes with learning systems Is you have a lot of inputs you have Probabilistic goo inside and then you Have outputs and and the problem is is Getting determinism out of probabilistic Anything right it's like a not it's a

Non-sequitur right like those two words Don't make sense Um and so when you look at where we most Likely will apply it is trained models That are then used and then we can Evaluate for determinism and then apply And I've done this in the past with my My last company so I know it can be done And you can get it safety certified to Do that so that's great Um but then also like there are lots of Areas where safety isn't involved and Then it can be super empowering you know Like all the things that we just talked About for like do this go here like Those types of business logic activities Could use things expert systems like Chat chpt to define the workflow and Help a non-expert user construct a Workflow that maybe they didn't have all The ideas for how they would get done You know like uh for you know metric or Business logic that they couldn't figure Out how to describe themselves but a an Expert system could help them describe Yeah or even something like go get the Red box right and then actually Understanding what that means Understanding to go get a box or Whatever and the other thing that I keep When you're talking about learning the Thing that brings to mind for me is that In the human learning process making Mistakes is a really big part okay right

But if you can do it in simulation right It's a better place to make mistakes I've never been to simulate my own Behavior but it's for robots that might Be the way to add to go well and that's That's a really comical one is because Typically most Learning Systems learn Through negative training examples Because humans learn a little bit Differently right when we see a dog as a Child and we learn that something is a Dog we also see a thousand examples of What is not a dog right but in in Training robotic systems we don't have All that luxury and so we'll say that's Not a dog or but it gets more Complicated when you talk about safety It's like okay don't hit things and What's the robot going to do learn to Hit something and you're like oh no you Shouldn't have hit that little kid like Right that's kind of an after effect Thing that you don't want to happen and So training some of these systems from a Safety perspective is hard when you're Using negative reinforcement yeah and so Does that mean that you have to be Particularly Um Uh Uh conservative about how you what you Allow a robot to do at any given time Right now yes I I think that that a lot Of a lot of the safety standards

Particularly are uh I generously I can say they're trying to Catch up Um but uh when you look at most safety Standards they really don't even Conceive probabilistic systems as part Of as part of the way that we even think About safety it's like it must do X and Then Y and then Z and if it doesn't do Those things it can never be safe kind Of thing right and so now the safety Standards Community is starting to Figure out like oh man this is reality So we got to figure this out because It's not going away well and I suppose With a human-shaped robot you end up in A situation very quickly where think of It thinking of it as a Cobalt becomes a Lot more it makes a lot more sense Where You Are a Human Being that's operating In a space alongside this robot in one Way or the other Um is that a use case that you guys are Exploring in the much longer term yes But I think there's an awful lot of work To do in industrial environments before We get to that point what is a much Longer term mean uh you know I'm not Well known for having the world's best Calibrated crystal ball all but you know I would say that's probably a post six To eight year kind of time frame but That again that's a guess Um I I guess that your bet is that while

You work towards getting there there's Still a large market for robots doing Robot things what human do well yeah I Mean as Melanie said right uh in the Time that she's been in the field there Are actually more jobs that need to be Filled over time so effectively uh Industrial Automation has made no Progress and has in fact lost ground Over the last 10 years yeah that makes Sense I think it's funny as well when You're talking about uh robot safety and Somehow you are oh there's a Venn Diagram diagram there between that and The OSHA regulations I'm like wait these Are robots aren't they meant to be able To be in a dangerous spaces yeah so so OSHA applies to worker safety right and If a robot is in its environment in a Person's environment that's a worker and So the robot has to comply with safety Related to that worker now if if workers Were to suddenly leave the entire Facility it still doesn't solve all the Problems because eventually someone has To go in and fix a robot right right Like if we get someday magically to the The point in which robots are fixing Robots that are fixing robots that are In its Turtles all the way down then Then we're good right but up until that Point OSHA is most likely going to be Involved because at some point someone Will be servicing a robot and if you

Look at the jobs that are like growing In the robot service robot technician Robot builder space they're they're Growing very fast because there's a lot Of robots to be built true yeah what was Left on the backlog like what what Didn't make it into digit where you're Where you wish you could push the Boundary or push the envelope for what Probably several hundred things Um let's go through the first 50 or Something Um you know something that we we did Very successfully with the Cassie robot Which was our early research platform uh It had very very highly Dynamic motion Uh and so we just were able to set a Guinness record for the fastest bipedal Robot uh running 100 meter dash right um Digit is probably capable of running but It's hard to tell and when you start to Add upper body complexity to it to Actually make it you know do useful work For customers right you can start to Dial back some of the core Dynamic Features that we had in earlier versions Of robots so I think there's always a Balancing act you know when we look at Yeah Uh when we look at things like battery Life there's always a question how much Battery do you put in that's a question Like how much food do you take Backpacking if you take a lot you walk

Really slow but you can stay out for a Long time if you take a little you can Run really fast but you'll have a Limited number of nights out so there's All the these trade-offs that are sort Of conflicting at all times requirements For the robot how far can does your Backpack Probably six to seven kilometers as a Guest I don't know we haven't tried it We we have actually a fun stat which is It's walked about 27 marathons worth of Steps uh since the promat uh show so That's pretty exciting we think that's Probably the farthest that bipedal Robots have ever walked that's because You keep walking it all the way to these Shows back and forth back and forth yeah Totally Um what do you wish uh had made it into The final robot Obviously a reality sensor I mean I I Think I think that the thing about any Roboticist is they're always going to Want more sensing like we want all the Bells all the whistles all the lidar but As Damien said like we have realistic Power constraints and the fact of the Matter is is like great we have all that Data but then we can't get it off the Robot because eventually it needs to go To a customer site and uh I don't know You know most customers don't like it When you have you know maybe 50 robots

And all of a sudden you've taken down All the bandwidth right so there is a Real trade-off that you have to make in Terms of great we have all this Perception capability but what can we Actually do with it in the long term and So balancing the Practical realities but It could make the robot super cool like But it does it's already super cool yeah Yeah from a geeky robotics perspective Yeah not from uh you know like okay well I think a big hand for a fun and Exciting robot And thank you so much for coming cheers [Music]