The robotics industry is on the brink of a transformative era, driven by advances in artificial intelligence and the integration of supervisory control systems. These systems enable robots to operate with partial autonomy, relying on human oversight to guide complex decision-making.
This approach balances the strengths of human intuition with the efficiency of automation, offering a practical pathway to realizing the long-awaited promise of intelligent robots. As technology evolves, robots are not only becoming more capable but also smarter learners, poised to reshape industries from logistics and space exploration to disaster response and beyond.
Robotics Roller Coaster
The robotics industry has seen boom and bust times. Following the first industrial-robot installation in 1962, enthusiasm for robotics blossomed. In the 1980s, investment in start-ups exploded. But the field imploded just as quickly by the end of the decade as robots failed to meet inflated expectations.
It became common knowledge that robots were incompetent and that software programming costs were grossly underestimated. The robotics industry suffered a long drought, with consolidation into a handful of companies that made incremental progress in robot performance over the next two decades. Programming remained difficult. Vision systems were expensive and unreliable except in carefully controlled environments, and hard to program and integrate with robots.
In the 2000s, robotics started to reemerge. Advances in computing power, in programming languages, and open-source code helped make robots more capable and easier to program. OpenCV — an open source computer-vision code library — enabled much faster, cheaper and more reliable vision-system development. The robot operating system (ROS), introduced in 2007, promoted code reuse for robotics, which helped robot start-up companies develop products remarkably faster and cheaper. ROS also offered an open system with which designers could easily interface new sensors and systems.
In parallel, the new AI was emerging, with deep learning, convolutional neural nets, and generative networks that were precursors of today’s large language model (LLM) systems. Early success was shown in using neural nets for machine vision, which rapidly became the uncontested superior method for image interpretation.
The Rise of Supervisory Control
In spite of robotics resurgence, the Fukushima nuclear disaster in 2011 demonstrated persistent limitations. Robots were unable to enter damaged reactors and perform tasks that would have significantly mitigated the disaster.
The Defense Advanced Research Projects Agency (DARPA) used this scenario as a theme to help accelerate robotics via the DARPA Robotics Challenge (2012-2015). Consistent with prior DARPA grand challenges, the tasks were extraordinarily difficult for robots, with the expectation of spurring rapid advances and commercial engagement (similar to the DARPA Urban Challenge in 2007, which helped spur industry involvement in autonomous vehicles).
In the DARPA Robotics Challenge, multiple teams ultimately succeeded in performing the difficult tasks. The approach in common was to use supervisory control. Under supervisory control, a system has partial autonomy, and a human operator provides higher-level instructions and corrections while the system is responsible for lower-level behaviors.
Supervisory control is now becoming ubiquitous in robotics. A very recent example is the Odysseus moon lander from Intuitive Machines. Remote operators in Houston were able to monitor progress and make orbital adjustments prior to landing. However, communications time delays precluded real-time control of the landing. Thus, Odysseus had to identify a landing site and perform controlled touchdown autonomously.
In the same spirit, NASA’s On-orbit Servicing, Assembly, and Manufacturing project (OSAM-1) will use supervisory control. The vehicle incorporates dual robot arms intended to perform relatively complex tasks in space. Remote control or teleoperation is impractical due to communications delays. Instead, terrestrial operators will give high-level commands and entrust the robots to carry out the incremental tasks autonomously.
By comparison, we are familiar with the levels of automation in self-driving cars. Level 0 offers the driver alerts, but the human operator is fully in control. At level 5, the vehicle is fully autonomous and no human driver is required. Between these levels, a human is involved at varying levels of abstraction.
GM-owned Cruise vehicles are largely self-driving, though remote human assistants engage as necessary. This approach has been exploited by Plus1 Robotics, which provides and supports robots for logistics/material-handling. Achieving 100% success in parcel handling is unachievable. But with (remote) human oversight and occasional assistance, these robots are highly effective. Plus1 recently reported performing an accumulated 1 billion picks.
From Supervisory Control to Intelligent Robots
In the ’80s, we naively assumed robots would be intelligent. The fallout was devastating for robotics. But a new AI has emerged. Image interpretation is dramatically more capable and easier to program than ever before.
Natural language interfaces (building on LLMs) offer simple, intuitive front ends to communicate with automation (including robots). Extending AI to understanding manipulation is ongoing research, but we can expect advances in the near future. Robots are starting to get the brains that were expected — but lacking — in the ’80s, and they will repair their former reputation.
But this will not happen in a single step. Robot intelligence will need to ratchet up, and supervisory control offers a pathway.
Robots that are imperfect nonetheless have proven to be valuable, as in the above examples with Odysseus, autonomous vehicles, and logistics robots. These examples incorporate modern advances in autonomy, and they provide valuable — albeit not foolproof — capability. With operator oversight, they amplify human productivity.
At the same time, the robots under supervisory control are trainees. Since the new AI learns from examples, each correction offered by a human supervisor is also an opportunity to learn. The supervised robots, already earning their keep, will have the capacity to become increasingly competent.
This is the process by which robots will incrementally build their intelligence, ultimately realizing and exceeding the inflated expectations of the past.
Far from the unrealistic expectations of the past, today’s robots step into roles where human oversight mitigates their imperfection and their capabilities improve through learning. This iterative growth promises a in which robots exceed our wildest dreams — not in a single leap but through steady, meaningful progress.
The pathway to smarter robots is here, and with it comes the potential to transform industries, solve critical challenges, and redefine the relationship between humans and technology.
By Wyatt Newman, Ph.D.
For more news, information, and analysis, visit our Disruptive Technology Channel.
