The robotics field is in a deployment obsession right now. That's the right obsession. Getting robots to actually work in real environments, with real variability, in front of real customers is the hard problem everyone should be focused on. Before starting MACH Dynamics, I got a firsthand look at one version of this problem: building Augmenti, a real-time sensor-driven coaching system for motorsports that had to read a human's context and surface the right information at the right moment. That experience pointed me toward a problem that doesn't have much industry attention yet.
Once robots are deployed, the next wall is the human.
the assumption baked into deployment
Most deployed robots today operate in one of two modes. Fully autonomous, where the human is out of the loop entirely. Or fully teleoperated, where the human is entirely in the loop and the robot is just the body. Real deployments are messier than either. In most real environments, the human is still present, still responsible, still making judgment calls the robot can't.
Supervisory control is the default, not the exception. A warehouse robot flags an edge case and a human decides. A surgical robot executes a movement and the surgeon maintains final authority. A manufacturing cobot pauses and waits for the operator. The human is always somewhere in the loop.
two versions of the problem
There are two versions of the human-robot interaction problem and they are not the same.
The first is explicit communication. The human tells the robot what to do. Voice commands, natural language, gesture. "Pick up the box." "Stop." "Hand me that." This is where most of the industry attention is right now, and it's getting solved. LLMs have made natural language robot control dramatically better in the past two years. Figure, 1X, and others have demos of robots responding to verbal instruction in real time. This version of the problem is hard but the path is clear.
The second version is implicit understanding. The robot infers what the human needs without being told. It reads context, state, timing, and acts accordingly. It knows when to hand off and when to stay out of the way. It knows what information to surface right now versus what would add noise. It knows the difference between a human who is in flow and one who needs intervention. No command required. The robot just knows.
The gap between those two is enormous. Almost no one in industry is working on the second one yet.
why the second version is hard
Three things make implicit understanding technically difficult.
Latency. In fast-paced environments, feedback that arrives 200ms too late is worse than no feedback. The system has to be fast enough that the human never notices the gap between what's happening and what the robot knows. This is a hard real-time systems problem, not a software problem.
Relevance filtering. Most sensor systems drown operators in data. The hard problem is knowing what matters right now for this specific human in this specific moment and surfacing only that. Everything else is noise that increases cognitive load and degrades performance. This requires the system to build a model of the human's current context, not just the environment.
Trust calibration. Humans overtrust and undertrust robots in ways that are both dangerous. A system that has been reliable for 1000 interactions and fails on the 1001st has broken something that's hard to rebuild. The robot has to earn appropriate trust incrementally, which means being right about when to act and when to defer, consistently.
where the research is
This is an active research area in academia, mostly under the label of shared autonomy. The two dominant approaches are policy blending, treating human input and robot action as two signals and arbitrating between them, and intent inference, where the robot builds a model of what the human wants and assists toward that goal. Intent inference is the harder and more interesting one. It breaks down in novel situations, which are exactly the situations where it matters most.
The most mature applied versions of this exist in surgical robotics (Intuitive Surgical's da Vinci system, where the surgeon maintains authority but the robot stabilizes and scales movement) and advanced cobots in manufacturing. Even in those contexts, the interaction is mostly explicit. The implicit layer is still largely a research problem.
why industry will have to care soon
As robots get more capable, the human-robot interaction problem gets harder, not easier. A robot that can do more things creates more ambiguous handoff points. More capability means more moments where the division of responsibility between human and robot is unclear. The more autonomous the robot, the more consequential every transition between human and machine control becomes.
The field will finish the deployment chapter. When it does, the next chapter is the human. The teams that have thought about this problem before it becomes urgent will have an advantage that's hard to replicate quickly.