Article Sidebar
Introduction
We began our journey in an academic context. Agility consisted of a small research team conducting experiments far away from the general public. In that environment, the safety of our systems was far from a priority, as the robots we made were for our own personal use. But the moment we began commercializing our solutions, safety began to occupy a central role.
We want to give you a look into what prioritizing safety has been like for our team, and explain how our approach is going to open us up to the largest addressable market.
Context matters
When we started discussions around what it would take to build a safe humanoid, we had to answer the following: safe for what? For where? And for whom?
Humanoids appear online in myriad contexts: in public demonstrations, tradeshow demos, in lab settings, and in real industrial deployments. Their seamless presentation on our screens can obscure a baseline truth: each of these environments requires much different safety considerations. At Agility, we intentionally decided to focus on that last sector first: industrial deployments in real factory settings. Therefore, building a safe robot for us meant investigating the specific guidelines and workflows we would encounter in those spaces, and working backwards from there.
Industrial safety
The rigorous standards in some of our first deployments, like our 2023 deployment at Amazon, taught us a great deal about what is required for industrial robotics safety. By the time we reached our third generation build of Digit, we only had a remote E-stop to provide functional safety. We improved upon that framework with the addition of an external safety controller that let us interface with the robot. But because it was located externally to the hardware, if there were communication issues with the robot it would fall - creating new risks not previously accounted for.
After more real world experience, it was time for some substantial updates. We added an onboard safety Programmable Logic Controller (PLC), essentially a dedicated industrial computer that controls a piece of machinery. This time, if the humanoid’s main control system had a software bug or made a bad decision, the safety PLC was there to step in. We also added category 1 stop functionality so that Digit could come to a controlled stop before power fully cut or if communication with the robot was lost, and we integrated a control pendant giving on-site customer personnel the ability for intervention. However, this was still not enough to guarantee the safety of people working around and with Digit. So each humanoid we’ve deployed in the field operates behind a set of physical safety barriers, colloquially known as a “workcell”, separating it from employees when it presents motion and stability risks.
As we get ready to unveil our fifth generation humanoid, which is also designed for industrial use cases, we are aiming to finally go beyond the barriers of the workcell. Our newest offering will be able to operate in the same space as people at a safe distance without relying on barriers – a designation called cooperative safety. This road to cooperative safety has required intensive investment in our own safety teams and has entailed forming partnerships with external developers, like NVIDIA, to achieve. More on that below.
Manufacturing, logistics, and distribution were the logical place for us to start because of all the assumptions we could make and validate about the people and spaces that we were going to interact with there. But of course, there are many more markets we believe humanoids will provide enormous value to. As we build solutions that expand our initial target markets, there will still be some restricted, controlled spaces like the back of house at hospitals or retail where trained employees also work. But as we start to move towards workflows like counter service where our humanoids may interact with the public, or are completely customer-facing like at a grocery store or in a home, the technological needs to guarantee safety will be far more advanced – and the evidence required to prove we’ve achieved a sufficient level of safety, even more significant.
This is such an important point to level-set on. When you see demos of robots operating out in the world around people and children that accidentally bump into or even strike a person, you can only arrive at one clear conclusion: these companies and OEMs have not prioritized their safety architectures. By starting in the most regulated and controlled environments, we are building a safety foundation of technology and evidence that will ladder up to increasingly complex and dynamic use cases.
Standards and regulators
Ours is still a burgeoning field, even though we’ve been hard at work for over a decade. That means that many of the standards that will govern our humanoids are still being written. Voluntary ISO/ANSI standards have historically split robots between industrial robots and what they termed “service robots,” the latter of which covered domestic or consumer robots like mobility aids. But with a human form factor and an increasingly sophisticated set of capabilities, humanoids are blurring that standard delineating line.
To help create the brightest possible future for our industry, we are contributing to and leading projects within ISO TC 299, Working Group 12 and ANSI/A3 R15.08. These groups of industry experts are actively building a new standard, ISO 25785-1 and a new technical report ANSI/A3 TR R15.108, that will cover safety requirements for dynamically stable industrial mobile robots. Building these as consensus standards instead of siloed-off corporate best practices is essential, because we’re able to get viewpoints from across the industry, gathering as much experience as possible. But the voluntary standards shaped by the ISO and ANSI are only one set of considerations humanoid providers have to keep in mind.
There are different and evolving regulatory requirements in different markets around the world . In addition to OSHA in the US, Europe currently has the Machinery Directive, which will soon be replaced by the EU Machinery Regulation 2023/1230 in early 2027. There are also bipartisan efforts in Congress advancing right now that advocate for similar regulation here in the US. We consider existing and emerging standards as well as requirements and guidance from government agencies from the earliest phases of our new designs, which will be essential as we serve customers across the globe.
AI safety and NVIDIA Halos
The last area we want to touch on is our efforts around AI safety.
First, our approach to AI safety standards. Here, the most relevant standard that we are contributing to is ISO/IEC TS 22440. To ensure that our humanoid demonstrates (and that the standards reflect) state of the art practices for the use of AI in safety, we believe there are two principles to be addressed. The first is quantifiable AI safety performance, including the detection of AI failures, which will ensure we measure the performance of AI models, identify the specific limitations and causes of their failures, and communicate and manage the residual risks. The second is monitoring and management of the deployed systems to ensure we detect and address underperformance of the models, new conditions not anticipated during development, and to address novel AI failure modes such as model drift. By checking the performance of our models and systems during deployment we can make sure that issues are identified and addressed effectively.
Second, are the AI software and hardware partnerships we’ve made that will allow us to pursue and achieve cooperative safety. The first announcement we’re excited to make is with NVIDIA who just introduced NVIDIA Halos for Robotics, expanding their safety system beyond autonomous vehicles to become the industry’s first full-stack, comprehensive safe compute platform for robotics and physical AI. The NVIDIA IGX Thor platform is a key part of the Digit V5 safety concept. We are collaborating with NVIDIA on how best to utilize this platform for humanoids and are using it in our forthcoming 5th-generation robot to run onboard safe human detection functions. This will be a major step forward for what our robots can do in the field, and will unlock a massive potential to scale beyond the workcell.
As we get closer to the unveiling of our 5th-generation humanoid, we will discuss further partnerships that are essential to building the safest humanoid ever brought to market.
Safety as strategy
Safety is sometimes framed as a constraint, a set of boxes to check or restrictions to deal with before you're allowed to operate. But we see the challenge to make a truly safe humanoid differently. Every safety milestone we achieve is a door that opens for us. Cooperative safety unlocks environments and use cases we couldn't enter before. Those in turn unlock the operational experience required to continuously evolve our safety technology and achieve the next tier of safety concepts and certification. That is how we unlock more use cases in industrial settings and expand into increasing levels of deployment: retail, healthcare, and eventually the home, all while adhering to the safety and compliance principles we’ve long held as essential here at Agility.
The addressable market at each level is dramatically larger than the one before it, and the companies that will capture those markets aren't the ones moving fastest today. They're the ones building the safety foundation that makes those markets accessible at all. That's the investment we're making with each update to our safety ecosystem and design, and it’s the one we’ve been making since we first left the academic environment over a decade ago.
