Why Reliability Is The Hardest Problem In Physical AI

Forbes

a day ago

Dr. Jeff Mahler: Co-Founder, Chief Technology Officer, Ambi Robotics; PhD in AI and Robotics from UC Berkeley.
getty
Imagine your morning commute. You exit the highway and tap the brakes, but nothing happens. The car won't slow down. You frantically search for a safe place to coast, heart pounding, hoping to avoid a crash.
Even after the brakes are repaired, would you trust that car again?
Trust, once broken, is hard to regain. When it comes to physical products like cars, appliances or robots, reliability is everything. It's how we come to count on them for our jobs, well-being or lives.
As with vehicles, reliability is critical to the success of AI-driven robots, from the supply chain to factories to our homes. While the stakes may not always be life-or-death, dependability still shapes how we trust robots, from delivering packages before the holidays to cleaning the house just in time for a dinner party.
Yet despite the massive potential of AI in the physical world, reliability remains a grand challenge for the field. Three key factors make this particularly hard and point to where solutions might emerge.
1. Not all failures are equal.
Digital AI products like ChatGPT make frequent mistakes, yet hundreds of millions of active users use them. The key difference is that these mistakes are usually of low consequence. Coding assistants might suggest a software API that doesn't exist, but this error will likely be caught early in testing. Such errors are annoying but permissible.
In contrast, if a robot AI makes a mistake, it can cause irreversible damage. The consequences range from breaking a beloved item at home to causing serious injuries.
In principle, physical AI could learn to avoid critical failures with sufficient training data. In practice, however, these failures can be extremely rare and may need to occur many times before AI learns to avoid them.
Today, we still don't know what it takes in terms of data, algorithms or computation to achieve high dependability with end-to-end robot foundation models. We have yet to see 99.9% reliability on a single task, let alone many.
Nonetheless, we can estimate that the data scale needed for reliable physical AI is immense because AI scaling laws show a diminishing performance with increased training data. The scale is likely orders of magnitude higher than for digital AI, which is already trained on internet-scale data.
The robot data gap is vast, and fundamentally new approaches may be needed to achieve industrial-grade reliability and avoid critical failures.
2. Failures can be hard to diagnose.
Another big difference between digital and physical AI is the ability to see how a failure occurred. When a chatbot makes a mistake, the correct answer can be provided directly. For robots, however, it can be difficult to observe the root causes of issues in the first place.
Limitations of hardware are one problem. A robot without body-wide tactile sensing may be unable to detect a slippery surface before dropping an item or unable to stop when backing into something behind it. The same can happen in the case of occlusions and missing data. If a robot can't sense the source of the error, it must compensate for these limitations—and all of this requires more data.
Long-time delays present another challenge. Picture a robot that sorts a package to the wrong location, sending it to the wrong van for delivery. The driver realizes the mistake when they see one item left behind at the end of the day. Now, the entire package history may need to be searched to find the source of the mistake. This might be possible in a warehouse, but in the home, the cause of failure may not be identified until the mistake happens many times.
To mitigate these issues, monitoring systems are hugely important. Sensors that can record the robot's actions, associate them with events and find anomalies can make it easier to determine the root cause of failure and make updates to the hardware, software or AI on the robot.
Observability is critical. The better that machines get at seeing the root cause of failure, the more reliable they will become.
3. There's no fallback plan.
For digital AI, the internet isn't just training data; it's also a knowledge base. When a chatbot realizes it doesn't know the answer to something, it can search through other data sources and summarize them. Entire products like Perplexity are based on this idea.
For physical AI, there's not always a ground truth to reference when planning actions in real-world scenarios like folding laundry. If a robot can't find the sheet corners, it's not likely to have success by falling back to classical computer vision.
This is why many practical AI robots use human intervention, either remote or in-person. For example, when a Waymo autonomous vehicle encounters an unfamiliar situation on the road, it can ask a human operator for additional information to understand its environment. However, it's not as clear how to intervene in every application.
When possible, a powerful solution is to use a hybrid AI robot planning system. The AI can be tightly scoped to specific decisions such as where to grasp an item, and traditional methods can be used to plan a path to reach that point. As noted above, this is limited and won't work in cases where there is no traditional method to solve the problem.
Intervention and fallback systems are key to ensuring reliability with commercial robots today and in the foreseeable future.
Conclusion
Despite rapid advances in digital GenAI, there's no obvious path to highly reliable physical AI. It isn't just a technical hurdle; it's the foundation for trust in intelligent machines. Solving it will require new approaches to data gathering, architectures for monitoring/interventions and systems thinking. As capabilities grow, however, so does momentum. The path is difficult, but the destination is worth it.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?