Robot autonomy has moved from getting a machine across a room to getting it to complete sequences of real tasks, and modern AI is the reason. Boston Dynamics vice president of software Matt Malchano, who has worked on autonomy for roughly 15 years, told Ars Technica that the definition of the word itself has shifted inside his own career. Physical Intelligence co-founder Sergey Levine, a computer scientist at the University of California Berkeley, frames the current frontier as reliable robot behavior in unstructured environments — a step beyond the millions of industrial and service robots already deployed on factory floors and inside warehouses.
The historical baseline is unflattering. In 1979, the Stanford Cart needed 5 hours to move 20 meters through an obstacle-filled room. The first bipedal robot that could walk without falling over was not built until 1996. Boston Dynamics itself only spun out of an MIT lab in 1992, and for most of the intervening decades, the industry's ambition was locked at point-A-to-point-B navigation.
“When I started maybe about 15 years ago, I led a project team that was focused on autonomy, but in that era, the goal of that team was to just get a robot to navigate from point A to point B.”— Matt Malchano, Vice President of Software, Boston Dynamics
Two AI shifts changed the trajectory. Reinforcement learning matured through the 2010s, giving robots a way to acquire specific motor skills through trial and error in simulation and in the real world. Then, in the 2020s, large foundation models trained on huge datasets began providing the prior knowledge robots need to interpret a scene, follow a language instruction, and recover from a mistake. Malchano said the combination unlocked the ability to imagine a robot understanding a task rather than just executing a scripted motion.
Key facts
- 01In 1979, the Stanford Cart needed 5 hours to move 20 meters through an obstacle course; the first self-balancing bipedal robot arrived in 1996.
- 02Boston Dynamics, spun out of MIT in 1992, now defines autonomy as sequences of tasks rather than point-A-to-point-B navigation.
- 03Physical Intelligence, co-founded by UC Berkeley's Sergey Levine, is building one general AI model to power many robot form factors, not a single humanoid.
- 04Millions of industrial and service robots already run in factories and warehouses on scripted motions — the research frontier is reliable work in unstructured environments.
- 05Reinforcement learning breakthroughs in the 2010s and foundation models in the 2020s are the two AI shifts driving the new robotics push.
The International Standards Organization defines autonomy in robotics as the ability to perform intended tasks based on current state and sensing, without human intervention. Levine breaks that down into levels. A factory arm repeating a fixed motion clears the basic bar. A robot handling a task reliably in a messy, changing environment — a warehouse aisle, a hospital room, a kitchen — is the level that is only now becoming a real research topic making its way into deployment.
Physical Intelligence is betting the payoff is not one universal humanoid. Levine argues the winning architecture is a general AI model that can drive many different robot bodies, each suited to its job. A small ceiling-mounted arm might make sense in a cramped New York City apartment; a heavy mobile platform makes more sense on a farm. Humanoids will exist, in Levine's view, but they will be one option among several.
Getting there requires what Levine calls step changes in technological complexity. A general-purpose robot has to perceive a novel environment, execute robust motor skills, recover from basic errors, follow human instructions, and generalize learned behaviors to situations it has not seen. Reinforcement learning refines skills through repetition — Levine's tennis-swing analogy — but only works if the robot has enough common sense to start.
That common sense is a data problem. Teleoperation rigs, where humans directly guide a robot through a task, generate high-quality training data but are slow and expensive. Physics-based simulations are cheaper but miss real-world messiness. Some companies are hiring gig workers to wear head-mounted cameras during household chores, feeding first-person video into world models that predict the consequences of actions. Each approach has a ceiling, and building world models remains computationally expensive.
Today's robots sit on an uncomfortable trade-off. Reinforcement learning can produce a system that performs one task at 99.99% reliability under narrow conditions. Broader training across teleoperation and video data yields robots that can attempt a wider variety of tasks but at lower reliability. Closing the gap — a robot that is extremely good at many things, not just one — is where the field's remaining research risk lives.
There are counterweights worth naming. The gap between a compelling demo video and a robot that earns its keep on a factory floor is still large, and the humanoid category in particular has attracted billions of dollars of investment on the strength of showreels rather than deployed unit economics. Agility Robotics has Digit units working in warehouses, and Boston Dynamics has commercial customers, but the broader promise of a general-purpose robot doing arbitrary household work remains a research goal, not a product. Safety, reliability under adversarial conditions, and the cost curve of teleoperation data collection are all unresolved.
For the AI market, the interesting move is architectural. If Levine is right that a single general model will power many robot form factors, the value in robotics concentrates in whoever builds that model — not necessarily in whoever ships the hardware. That mirrors the pattern already visible in software, where foundation-model providers capture disproportionate value versus the applications built on top. Investors betting on humanoid hardware companies without a model story attached may be buying the wrong layer of the stack.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
