Industrial robots are highly capable within their domains of expertise, the environments and scenarios that they know. Take them out of their comfort zone, and problems start occurring, performance degrades or stops, or even worse, damage occurs.
For modern machine-learning-based robots, that domain is defined by the training data, so when something changes, the usual course is to collect new data and retrain a model. As well, engineers may adjust parameters or rewrite exception-handling logic. If new hardware needs to be added to the system to handle the novelty, the integration work can become even more involved. All of these steps require that the system be taken offline, impacting productivity.
If you’ve been following logistics automation, you’ve probably seen the demos from Figure, XSquare, as well as Plus One running autonomous induction lines that handle a wide variety of package and parcel types. The setups differ, but the underlying approach is shared, and so is the failure mode.
For example, imagine a robot that has been working on a warehouse induction line for hours, picking and placing packages. Suddenly, something unfamiliar arrives, like a plastic bag instead of a cardboard box, a loose apple or a package format the robot wasn’t trained on. This is where the problem begins, and the data collection and retraining described above starts.
This dependence on retraining has become one of the biggest constraints on warehouse automation. By nature, warehouses are constantly fluctuating with different product mixes, seasonal peaks and new equipment. Operators are always looking for ways to increase throughput in the face of this variability, and they need automation to evolve with them instead of pausing for a retraining cycle every time something changes.
Some believe that a sufficiently general robot, a universal robot, is the answer. That for such a system, nothing would be out of scope, so it would never need to be retrained.
Unfortunately, we don’t have such a robot today, so instead we turn to adaptive robotics: systems that continuously learn from experience, human guidance, and their own performance to ultimately eliminate the need for extended downtime.
A brief history of now
Until recently, industrial automation worked best when engineers could define a problem in advance and program robots to handle it. If the equipment involved, the sequence of steps, and even the dimensions of the product are known, developers can create algorithms to optimize
system performance.
Machine learning expanded those capabilities with things like imitation learning (IL), training a system from examples instead of relying only on manually coded rules. While the idea has been around for a while, advances in model complexity, training systems and compute have recently moved it from the lab to the workplace. IL lets a system absorb the specific variability of a target domain without having to model it — unless you’re doing sim2real training, in which case, an explicit model is still required.
Either way, the learning is front-loaded with teams collecting data, automating it, training the model, then deploying it. When conditions change, at least some part of that process needs to be repeated. This creates friction, especially when it comes to automation in environments with many variables, or when a new application or product domain is added. At that point, the whole process needs to begin again.
This operational burden could eventually be reduced by robotics foundation models and embodied AI that learns to perceive and act through physical systems. The technology is quickly advancing, but industrial operations have specific requirements around speed, reliability and
uptime. Unfortunately, emerging general-purpose models don’t consistently meet these requirements yet.
Given this reality, how can operators get more adaptable automation using the warehouses and equipment they already have?
Teach the robot when the unexpected happens
The best approach is to teach the robots using human-in-the-loop (HITL) AI. When a robot runs into something it doesn’t know how to handle, a person steps in to tell it what to do, and the episode turns into an example the system can learn from.
A version of this approach is already working today in exception handling. When a robot encounters a scenario it can’t manage, a remote operator steps in so that the line can keep moving. The data captured at that moment feeds straight into the learning process, eliminating
the need for a separate team to label every example manually.
The next evolution of this intervention is to extend robot learning beyond just noticing the error and into taking action, recovery behavior, and system-level decisions.
Human demonstrations are a good starting point because they show the robot one viable way to solve the issue. But a system trained this way is only as good as the humans it learned from. Maybe a little faster or more accurate, but fundamentally bound to the demonstration. In fact, at times it needs multiple demonstrators to perform the task the same way or it gets confused.
This is where reinforcement learning (RL) comes in. The robot starts with what was demonstrated, then explores alternative solutions and improves based on the results it gets.
Warehouses face an enormous amount of possible situations and exceptions, and asking an RL system to find a solution through trial and error alone can take way too long. Human guidance narrows down that search. The robot starts with a solution that works, refines it through
experience and asks for help when it encounters something new.
Learning what can’t be predicted
One of the hardest problems in industrial automation is something called unknown scenario variability. If a robot is designed to handle cardboard boxes from one inch to one foot in size, it is ready to handle any variation between those two size parameters, even though it doesn’t know exactly what is coming next.
But if tomorrow’s item is a plastic bag or an apple, that is a format nobody anticipated when the system was designed. Engineers can’t pre-train the system for every future object because they won’t necessarily know what those objects might be.
Adaptive robotics changes the goals from anticipating every possibility to giving the system a mechanism for learning when something unexpected happens. Humans handle the novel cases, the system learns from the intervention and subsequent situations can become a part of the robot’s autonomous capabilities over time. It’s a learning loop that can last for the lifetime of the deployment.
What if the robot itself keeps changing?
The same adaptive learning can apply to hardware too. Robotics systems are usually developed around fixed embodiment. This means the cameras, arms, conveyors, grippers and other physical components are known in advance. Their overall performance depends on how
effectively they coordinate.
It turns out that hardware change is just as challenging as task change, just at a different level. When new cameras, arms or conveyors are added, they function as novel input and output dimensions, and the system has to learn the mapping between them that helps it accomplish its tasks. Rather than the scenario changing, the robot’s capabilities have changed, but the goal is still the same.
Coordination follows the same pattern. For example, a robot running at full speed may overwhelm equipment downstream and create congestion. Another machine may need to slow down, change task order or temporarily redirect work. The most useful optimization target then becomes throughput across the entire operation.
This is where orchestration comes in. The interesting part is that inter-system optimization is really the same problem as intra-system optimization, just at a higher level of abstraction. Inside an induction cell, optimization means coordinating exact hardware movements. At the system level, it’s the same coordination problem, only now it’s applied to parcel locations and picking targets.
As fleets become more diverse, facilities need an intelligence layer that can allocate work, manage resources and coordinate different types of equipment. Connected systems can even share what they are learning with other robots, improving performance across the whole fleet.
The impact of adaptive robotics on automation
The ability to integrate and coordinate new automation within an existing system could also make warehouse automation easier to scale incrementally.
Most facilities aren’t going to automate everything at once because budgets, timelines and risk generally push companies toward specific bottlenecks first. They may do it in stages as needs and economics justify each investment. For example, beginning with automating induction, then palletizing, followed by material movement.
New automation can become part of an existing operating environment without requiring every existing system to be redesigned around it. The warehouse may automate one layer at a time while optimizing how those layers work together.
This could change the calculation behind automation investments by factoring in deployment speed, integration effort, and the cost of adapting systems alongside traditional business case metrics like throughput and labor savings. A system that can absorb new tasks, equipment, and operating conditions may have a lot more value than a robot that performs well on Day 1.
Resilience is the next performance metric
Industrial robotics is moving from designing automation around a fixed definition of the job to designing systems around change itself. As warehouses add more autonomous equipment, the focus shifts from making individual machines smarter to making the entire operation smarter. While it’s impossible to anticipate every case in advance, resilience and the ability to keep learning as the operation changes may ultimately become as important as speed, if not more.