1. A Convenience Store Robot Controlled by a Person in a Distant Office
The company that built the robot is Telexistence. The operator sat in an office in Toranomon, using VR equipment to move the robot in the distant store. The first task was arranging plastic-bottle drinks in the back room of the store[1].
The stated goals were threefold: addressing labor shortages, creating new jobs where people can work from anywhere, and reducing physical contact during the pandemic. The system was designed so one operator could manage multiple stores [1]. This combination—a robot doing the physical work, guided by a human—raises a key question: if robots are meant to take over some of people’s tasks, why is a person still involved?
2. Why Japan Needs Robots: The Working-Age Population Has Shrunk by About 13 Million
To understand the need for robots, we must look at the numbers. According to the Cabinet Office, the population of people aged 15 to 64, who form the core of the workforce, peaked at about 87.16 million in 1995. By October 2024, this number had dropped to about 73.73 million [2]. That is a decrease of roughly 13 million people over about 30 years.
Looking ahead, the ratio of working-age people to those over 65 is projected to fall to 1.3 people per elderly person by 2070 [2]. In a specific store, the burden is visible. At the FamilyMart METI branch, staff used to arrange about 1,000 bottles a day. This work took place in a cold environment, which is physically hard on the body [3]. The government also drafted a strategy in March 2026 to promote AI robots in 16 fields, including manufacturing and logistics [4].
3. Why Industrial Robots Handle Single Tasks Instead of Whole Jobs
Japan is a leader in industrial robots. In 2024, about 450,000 industrial robots were active in the country, ranking second in the world after China [5]. Factory robots excel because the environment is controlled: parts always arrive in the same place and orientation.
Convenience stores are different. Jin Tomioka, who was then Telexistence’s CEO, noted that a Lawson store has about 2,000 types of drinks and prepared foods, with around 200 different shapes [6]. Outside the factory, the world is full of variables and randomness. In August 2022, Telexistence began deploying the "TX SCARA" robot to 300 FamilyMart stores [7]. An AI called "GORDON" analyzes shelf data and sales trends to decide which products to restock first [8].
However, if the AI encounters something unexpected and fails, the system switches to remote control mode. A human then uses VR to fix the issue [3][7]. By January 2025, deployment was still progressing across about 300 stores [9]. The robot does not do the whole job. It handles repetitive tasks, while humans handle the unexpected.
4. How Robot AI Learns by Imitating Human Movements
One way AI learns is through "imitation learning." This involves recording human movements and using that data to adjust the robot’s actions. In one study, researchers used remote control to demonstrate six fine motor tasks, such as threading a cable tie or opening a condiment cup. With only 10 minutes of demonstration data, the success rate reached 80 to 90 percent [10].
Another research team used 13 robots over 17 months to record more than 130,000 episodes of 700+ tasks. For familiar tasks, the success rate was 97 percent. For new tasks, it dropped to 76 percent [11]. Telexistence announced in 2025 that they planned to start a service providing robot motion data to other robot developers and researchers in January 2026[12]. Here, the remote-control technology used in the store is being extended to make demonstrations for robots to learn from.
5. How Robot AI Learns by Practicing in Simulation Thousands of Times
The second method is "reinforcement learning." The AI tries different actions and learns to give higher scores to movements that work well. It repeats this process in large quantities. In a famous experiment, a robot hand learned to solve a Rubik’s Cube. Because real-world practice is slow, the AI practiced in a computer simulation. The paper states this equated to about 13,000 years of experience [13].
To make the simulation realistic, researchers used "automatic domain randomization." They automatically varied the cube’s size, friction, and gravity during training. This is like practicing soccer with different ball sizes and slippery fields. After this training, the robot hand could solve real cubes, even while wearing rubber gloves or being poked by a plush toy [14]. However, the success rate was 60 percent for a standard scramble and only 20 percent for the hardest scramble [13].
6. Why Simple Human Actions Are Hard for Robots (Moravec's Paradox)
Robotics researcher Hans Moravec pointed out a paradox in his 1988 book Mind Children. It is relatively easy for computers to play checkers or do well on IQ tests. But it is very difficult, or nearly impossible, to give a computer the perceptual and motor skills of a one-year-old child [15].
Moravec argued that human sensory and motor skills evolved over a billion years. Reasoning is just a thin layer on top of this deep foundation. This is known as "Moravec's Paradox." We see this in the convenience store robots too. In tests with changed backgrounds, success rates dropped to 59 percent [11]. The hardest Rubik’s Cube scramble had a 20 percent success rate [13].
However, this view has critics. AI researcher Arvind Narayanan argues that we may be focusing only on striking examples. He suggests that reasoning may also require common sense [16]. Numbers from a few experiments cannot prove that the paradox is right.In the store, the system works by combining AI for routine tasks and humans for the rest.
7. How Home Robot Vacuums Began Without Maps
Home robots, like vacuum cleaners, have also evolved. The first Roomba, released in 2002, had no map. It moved almost randomly and changed direction when it hit an obstacle [17]. In 2015, the Roomba 980 introduced camera-based mapping for consumer products, allowing it to see and map the home [18].
In Japan, iRobot Japan announced in 2024 that their robot vacuums had reached a 10 percent household penetration rate by the end of 2023 [19]. This success may come from the fact that cleaning is a repetitive task on a consistent floor. In one research project, a robot learned tasks such as stir-frying shrimp and riding an elevator from 50 demonstrations per task [20]. However, these are research results, not products sold to families.
8. Three Questions to Ask When Reading News About Robots
If you read news about robots helping with labor shortages, ask three questions. First, which specific "task" is the robot handling? Second, who helps when it fails, and how? Third, how much did the AI practice, and where?
Whether an article or product answers all three tells you a lot about it.You can watch videos of the ALOHA project’s robots performing tasks [10]. You can also compare Telexistence’s announcements from 2020, 2022, and 2025 to see how their approach changed [1][7][12]. This example shows how this store’s automation combines work done by robots with help from people.