Article Sidebar
An enticing vision
Bringing robots into your home has always been an exciting prospect. Helping us with tedious tasks, offloading everyday chores, giving us more time to focus on the pursuits and people that we love. Many of us use machines and robots on a daily basis, but they have become so commonplace that we no longer see them for the complex pieces of engineering that they once were – a Roomba in the dining room, a washing machine churning away, the microwave heating up some popcorn. They operate in the background with the assumption that they will work at a moment’s notice with predictable results, and crucially, that they will be safe.
Since humanoid robots entered the market, the promise of having a new tireless partner that could help us with even more complicated tasks has been too enticing to ignore. What if a robot could do the entire laundry cycle for you, folding and all? What if a humanoid could help you landscape your yard? Pick up the mess your toddler made? Change the oil in your car? We’ve all seen the cappuccino pouring demos on YouTube or the humanoid wearing a basketball jersey like it’s just one of us. With a human form factor and a brain powered by AI, it can feel like we’re right on the cusp of science fiction becoming reality.
However, the truth on the ground is more complex than those fanciful scenarios might suggest. There are three main barriers preventing this vision from being realized today: humanoid capability, cost, and safety. Agility is working hard to tackle these challenges so that we can deliver a humanoid for the home when it’s responsible to do so. What follows is the realistic path as we see it.
Capability
As we’ve written about elsewhere, the current boom in AI technology has fueled the already prevalent misconception that a humanoid will soon arrive at your doorstep. However, training LLMs to operate effectively and training the software systems controller operating a humanoid robot are vastly different challenges.
The details are in the data. LLMs were able to capitalize on and be built on top of the massive wealth of written text data that existed on the internet. Virtually the entire human corpus existed for free online. Firms not only trained their models on patterns of speech, but also on mathematics, coding, and imagery as well. The resulting output from these models is staggering.
However, no corresponding reservoir of locomotion data exists for robots to train and act in accordance to. As much as we’d like the reality to be that we simply run Digit on one of the enterprise LLM models and it could suddenly walk and talk exactly like humans do, that isn’t the reality. Getting a piece of mechanical engineering to move autonomously in physical space, i.e. embodied AI, is far more complicated. It requires the collection and coordination of data sets that are quite different in nature – including the forces involved in a humanoid's movement, different lighting conditions, degrees of freedom, joint limits, velocities, safety boundaries, and contact dynamics, to name just a few.
While we were able to construct robots without artificial intelligence models, we have since been busy building increasingly sophisticated foundation models to help our humanoids move. We’re generating the types of data listed above to build our controllers using learning from demonstration as well as reinforcement training methods in powerful simulation platforms. But there is much work still to be done. In other words, don’t let the teleoperated humanoid fool you. You might see a robot operating in a public space crowded with people, giving you the perception that humanoids can already autonomously perform these complex behaviors, but it’s likely controlled by a person somewhere out of sight. This isn’t a fully autonomous robot that could scale in a helpful way. It’s a deceptively disguised trick.
But we already have humanoids operating in industrial environments today at Schaeffler, Toyota among others. Haven’t we already arrived? Unfortunately, the home is one of the most chaotic places a robot can operate with conditions that change on a day by day, or even minute by minute, basis. This is not a robot moving from point A to point B with static conditions like you would experience inside of a highly regulated industrial environment, or inside of a closed system like a work cell. Further, the array of tasks we’d hope for a humanoid to handle inside a home is much larger than those in a repetitive labor environment. Simply put, we still need much more data – each tranche specific to the unique actions the robot would be called to take. By taking the data from our real world deployments (and from the simulations we’re running back in our own facility) and learning from them, we are on the path to safely develop more intricate capabilities. But the industry must be patient and not make unrealistic promises.
It’s helpful to take another popular example beyond LLMs to demonstrate the enormity of what we mean. See: the world of self-driving cars. The companies building self-driving vehicles are collecting data every day on the streets around us with massive fleets of vehicles, sometimes with a human behind the wheel, or with a remote operator on stand-by to intervene in an emergency. When they’ve gathered sufficient data and adequately trained their systems, they will be ready to move onto the next step of having fully autonomous cars that we can rely on. But the range of motion for a car is dramatically more narrow than a humanoid robot in the wider world, even though the type of street these cars drive on varies quite a bit. Cars must stay on the road, they can go forwards and backwards, left and right. Compare that to a general purpose humanoid operating in the public that must be able to go anywhere we can go – help with the large array of tasks that we do. The amount of scenarios we must plan and train for is exponentially higher.
Humanoid capability currently lags what’s required to enter the home, but we will achieve the needed skills by training in controlled environments first where we can be assured of safety. We will build the reservoir of data we’ve outlined and eventually be capable of building a capable autonomous robot that can scale.
Cost
Any realistic conversation about building a humanoid intended for the home must include the issue of cost. At the customer level, individual families are very cost-sensitive. Even if we assume these robots are useful in as many areas as we believe than can be, they still must be affordable and cost no more than a family car in order to be a widely adopted product category. But in their current state, capable humanoids are much more expensive to produce. How can we scale to reduce the cost?
Agility decided to first tackle use cases that are more realistic with our current safe and reliable capabilities – moving totes, palletizing and depalletizing. There are millions of unfilled jobs in the manufacturing and logistics spaces where these simple workflows are high impact and low complexity when compared to working in the home. As we continue to scale in these industries and are able to ship and build more humanoids, the cost will come down – all the while maintaining a sustainable business strategy in the process. Instead of going after the unrealistic goal first – heading straight to the home – we have chosen this more deliberate path. We believe it’s the only way to responsibly and safely deliver.
The best pathway to home leads us through many other environments first. Industries that Agility will continue to bring value to even when we’re eventually ready to deploy inside of a home context. As we move towards our first cooperatively safe robot, a designation that means a humanoid can operate outside of a work cell, we are prepared to open the aperture on the use cases and industries that we can reasonably access. Always gathering data along the way, always developing a finer more complex set of skills.
Safety
The final, and arguably the most important, barrier preventing us from entering the home market today is safety. That doesn’t mean an autonomous robot that works 95% of the time. That still leaves a margin of error that is unacceptable. It means delivering a robot that is as reliable as the commonplace technology that surrounds us that has faded into the background. Safety doesn’t mean what a humanoid’s intent is, but the statistical calculation of risk of human injury, validated by 3rd parties. That means conducting formal Hazard and Risk Assessments (HARAs), and then planning the mitigation for each and every risk.
We are still working out the kinks here. That means accidents. That means the robots fall. This is acceptable if the robot is working inside of a confined work cell, behind safety barriers separating them from human workers. But the home context is one where we must be uncompromising. We cannot tolerate an injury caused by a humanoid delivered before it was ready to be deployed. We have to control for pinch points when moving a turned-off robot, or having hands on it when it is actively moving. We need to account for robot judgment errors when it’s deciding to move something heavy or hot. When robots are allowed to physically interact with and touch people, how does it sense what’s a person and what isn’t in a given environment? What dictates when and where can they be touched? How does it calculate the appropriate forces required? A set of standards by which we can judge these robots is essential – not just the opinion of a private company.
Official regulations are the best safety strategies, documented and shared as an international standard via ISO or similar. With these in place, everyone in our industry can critique, refine, upgrade, debug, and familiarize themselves with the same strategy. With these in place, an insurance company will know whether a new robot entering the market is compliant or not. At Agility, we’re taking an active part in bringing these standards to life. We helped write the forthcoming Type-C safety standard ISO 25785-1 for industrial mobile robots with actively controlled stability, and contributed to an upcoming ANSI/A3 technical report for dynamically stable robots. All steps getting us closer to regulated humanoids intended for home use.
A promising, realistic vision
A future of abundance awaits us, where a tireless humanoid partner can help with a wide array of work and laborious chores. However, to arrive at that day the robotics industry needs to approach the challenge in a responsible way, garnering trust from our partners and the people who rely on our robots to be useful and reliable. The robots Agility makes today for heavy industrial tasks will eventually be complemented by lighter, smaller, more dexterous offerings that are more appropriately scaled to work in a home, or in a grocery aisle, or assisting with electrical work in the tight crawl space in your attic. And we are confident that that day will come – but first we must finish the foundation we’ve been diligently laying.
