Robotaxis without drivers already roam city streets, delivery drones drop packages at front doors, and the idea of general purpose robots helping humans at work or home no longer seems like science fiction. But turning that vision into reality depends on giving machines a much higher degree of autonomy, powered by modern artificial intelligence. That ambition has drawn billions in investment and turned many researchers into startup founders.
From navigation to autonomous task execution
Matt Malchano, vice president of software at Boston Dynamics, recalls leading a team focused on autonomy about 15 years ago. Back then, the goal was simply getting a robot to move from one point to another. Now the conversation has shifted to a vast range of tasks a robot might handle on its own. The path to general purpose autonomy once seemed impossible when labs were still struggling with basic navigation and balance. In 1979 the Stanford Cart needed five hours to travel 20 meters through a cluttered room. The first bipedal robot that could walk without falling emerged in 1996.
Robot autonomy has always been a moving target, Malchano says. The International Standards Organization defines it as the ability to perform intended tasks based on current state and sensing, without human intervention. Recent AI breakthroughs, especially reinforcement learning in the 2010s and large foundation models in the 2020s, have unlocked new possibilities. Multiple labs and companies are now racing to build robots that can handle a wide variety of tasks in complex, unpredictable environments.
Current robots and real world deployments
Not every useful robot needs to be humanoid. Boston Dynamics already sells its four legged Spot robot for autonomous inspections in hazardous settings such as substations and culvert pipes. Spot learned to walk on slippery floors through additional reinforcement training, adapting its gait similarly to how a human crosses ice. The company also supplies Stretch, a wheeled robot with a large arm, for warehouse work and package handling at logistics firms like DHL.
Boston Dynamics is now scaling production of the all electric Atlas humanoid robot. Atlas is being trained at a Hyundai Motor Group facility in Georgia, with the goal of deploying it at the automaker’s EV plant by 2028. The company aims to produce 30,000 humanoid robots per year by then, though it remains to be seen whether workplaces can absorb that many units profitably.
Challenges on the road to general purpose robots
Sergey Levine, a computer scientist at UC Berkeley and cofounder of Physical Intelligence, says the next level of autonomy involves performing tasks reliably in unstructured environments. His startup is working on a general AI model that could power many different robot forms rather than one super humanoid. A small robotic arm might suit a tiny apartment, while a large machine could work on a farm.
Developing such robots requires handling complex perception, robust motor skills, and the ability to recover from mistakes. Levine explains that researchers combine reinforcement learning with pretrained foundation models. Reinforcement learning is like practicing a tennis swing many times, but you first need basic common sense. Foundation models trained on images and text provide that world knowledge.
Still, a data gap remains. Collecting physical training data through teleoperation or real world trials is expensive and slow. World models trained on video can help but are computationally heavy and may not capture all real world physics. Current methods can produce robots that excel at a single task under specific conditions or handle many tasks with mediocre reliability. Levine wants the combination: extreme competence across all tasks, which remains at the frontier of research.
Safety is also a make or break issue. As robots enter human environments, failures could cause harm or damage. Researchers and companies are investing heavily in robust testing and redundant systems. The path to general purpose home robots is still long, but specialized autonomous machines are already proving their value in controlled industrial spaces. For the latest on how these technologies evolve, check out the latest AI news on Mylistingo.
