Toyota just rolled out a 7-foot-2 robot that drains three-pointers. Cute party trick, right? Except this one, called CUE7, ditched the old-school “program every move” approach and learned to shoot the way humans do: by missing a ton, adjusting, and getting better.
That’s the real headline. CUE7 isn’t running a pre-baked script anymore. Toyota says it’s using reinforcement learning, an AI method that tweaks behavior based on success and failure, so the robot can refine its mechanics shot after shot. And Toyota isn’t building a novelty act for halftime shows. They’re using basketball as a stress test for factory robots that need to cope with the messy reality of manufacturing.
CUE7 weighs about 163 pounds and stands 7 feet 2 inches tall. Same specialty as earlier versions: three-point shooting. The difference is what’s happening under the hood. Instead of executing a lab-perfect sequence of joint angles, it analyzes each attempt, figures out what went wrong (or right), and adjusts things like release angle, force, and arm position.
Reinforcement learning: make the shot, get rewarded; brick it, pay the price
The concept is brutally simple. Swish a shot: reward points. Miss: penalties. Over time, the algorithm hunts for the set of movements that boosts the make rate.
Toyota says CUE7 started out rough, under 20% accuracy in early tests in September 2025. Three months later, at its official Tokyo presentation, it was hitting 85% of its three-pointers from five different spots.
That’s the selling point of reinforcement learning: it can improve through experience without someone spoon-feeding it thousands of labeled examples. But the physical world bites back. Toyota’s engineers had to deal with wear and tear, CUE7’s joint motors can handle only about 50,000 cycles of heavy use before maintenance.
So they cheated the smart way: simulation. The Toyota Research Institute built a physics simulator that mimics the biomechanics of shooting, letting the AI run the equivalent of 100,000 virtual shot attempts per day. Let the robot “miss” in software, save the hardware for the shots that matter.
Toyota’s real target: factory robots that don’t freak out when parts aren’t perfect
Toyota’s pitch is straight industrial. Today’s assembly-line robots are obedient to a fault: they repeat the same motion endlessly, but only if the world stays perfectly controlled. Real factories aren’t perfectly controlled. Parts shift. Materials flex. Humidity changes. Stuff arrives slightly warped. And traditional robots can stall when reality drifts outside their narrow tolerances.
Akifumi Tamaoki, who runs Toyota’s robotics program, says CUE7 is a proving ground: “CUE7 allows us to validate the adaptation algorithms we’ll integrate into our assembly robots as early as 2027. The ability to automatically adjust movements when parts are slightly deformed or positions vary represents a considerable productivity gain.”
The juiciest example Toyota points to is body assembly. Sheet-metal tolerances can vary by a few tenths of a millimeter depending on temperature and humidity. A conventional robot may stop when a part falls outside expected parameters. A reinforcement-learning robot could adjust its grip in real time, Toyota projects that could cut production stoppages by 40%.
Honda and Boston Dynamics aren’t waiting around
Toyota isn’t alone in the robot arms race.
Honda is reportedly teeing up an “ASIMO-X” reveal for March 2026, also leaning on reinforcement learning, but with a broader ambition. Toyota’s robot is a specialist (shoot threes, get better at shooting threes). Honda wants a generalist that learns adaptive walking, object handling, and verbal interaction in parallel.
Boston Dynamics is taking a different route with “Atlas Pro,” expected to go commercial in late 2026. Their bet: imitation learning. The robot watches humans do complex tasks, copies them, then uses AI to optimize the motions.
And yes, the lack of a single dominant approach is a tell. The industry’s still arguing about what works best. Reinforcement learning can be great when you’re optimizing a repeatable skill. But throw a robot into a totally unfamiliar situation and imitation learning can have the edge, because it starts from human examples instead of trial-and-error flailing.
The buzzkill: this brain burns electricity
Here’s the part factory managers actually care about: power draw. Toyota says CUE7 pulls about 2.4 kW while operating, roughly double a “classic” industrial robot doing an equivalent task.
That’s not a rounding error if you’re talking about deploying fleets of these things. Toyota says it’s exploring ways to cut the energy hit: specialized neuromorphic chips, leaner learning algorithms, and hybrid setups that alternate between “learning mode” and efficient execution.
Because the math is ugly: if adaptive robots jack up electricity use across an auto plant, the productivity gains start fighting the company’s emissions targets and energy bills. Toyota knows it can’t sell “smarter robots” that also quietly spike the carbon footprint.
Toyota plans pilot deployments in three Japanese plants starting January 2027, with an internal goal of reaching “energy neutrality” through algorithm optimization.
