A convenience-store trial run on Amazon's cloud shows what the next generation of retail robots can do, and how much of their training still comes from people
Amazon Web Services spent part of this year helping a Tokyo robot maker test whether a machine at a shop counter can learn to think one step ahead.
The robot maker is Telexistence, which says it has deployed more than 300 robots with major Japanese convenience-store chains. Its machines restock drinks, and its newer humanoids are being trained to scan and bag shopping at the register. Most of what they know came from people. Human operators steer the robots remotely through real tasks, and those sessions become the training data.
In a write-up AWS published in September, Telexistence and scientists from the AWS Generative AI Innovation Center described fitting one of NVIDIA's newest models onto that setup. The model, DreamZero, is a world action model. Before it moves, it predicts what its cameras should see a moment later and plans its arm movements to fit. Telexistence's current robots respond to what is in front of them. DreamZero tries to anticipate.
Amazon supplied the computing and the expertise. The model was fine-tuned and tested on AWS servers running NVIDIA chips, through the Physical AI Fellowship that AWS runs with NVIDIA and MassRobotics. Telexistence says it was one of nine companies picked for the 2026 cohort and the first from Japan.
The results were uneven, and the engineers said so. In a simulator, the best version of the model completed a block-picking task 45% of the time. Trained on recordings from Telexistence's physical robot and then tested in the same simulator, it succeeded zero times. The team blamed mismatches between the real robot and its digital copy, along with noise in the recordings. On the fastest hardware tested, each step took about 583 milliseconds, which the team said was close to what real-time control requires. AWS notes the figures come from internal testing.
Genki Sano, Telexistence's co-founder and chief technology officer, said the work showed "where DreamZero helps, where the real-to-sim gap bites."
Read next: Amazon launches worker robot that takes conversational instructions
Who teaches the robots
Telexistence has pitched remote operation as a staffing idea for years. When it launched its Model-T shelf-stacking robot, it said store staff could work from anywhere with an internet connection. The DreamZero trial shows the other half of that arrangement: how much the robots depend on what people feed them.
The team went through 183 recorded demonstrations of the block-picking task. Human reviewers threw out 91, almost half. In some, the robot sat idle for long stretches. In others, the operator picked up a block of the wrong color, or someone's hand drifted into the camera's view. The engineers also built an AI reviewer to do the screening automatically. It flagged 72, or 39.3%, close enough to work as a first filter, they said, though it had trouble telling blue blocks from purple ones.
Read as an HR document, that is a job description. Someone has to perform a task cleanly and consistently for a camera, and someone else has to check the footage. Neither role fits neatly into a retail org chart today. Both will need standards, training and a pay rate.
HRD has reported that when staff are filmed to train robots, employers should handle it as workplace monitoring, and agree in writing who owns the footage, whether people are paid for it and whether it can be used in performance reviews.
Read next: How close are robots to replacing human workers?
Why the pressure is building
Japan is the test market because it is running out of workers. The Recruit Works Institute, a think tank run by Japanese employment company Recruit, projected in 2023 that the country would be about 3.41 million workers short in 2030 and more than 11 million short by 2040.
Seven-Eleven Japan, which has more than 20,000 stores, agreed in September 2025 to develop a humanoid with Telexistence. The robot, called Astra, is due in stores in 2029. The companies said the aim is to "transform the very role of employees in convenience stores," handing routine work to robots so staff can focus on service only people can provide.
Canada's federal outlook is calmer. Employment and Social Development Canada counted about 349,600 cashiers in 2023 and expects labour demand and supply for the job to stay broadly in line through 2033. Its three-year outlook is weaker in the biggest markets, where it rates prospects as limited in Ontario, Quebec, British Columbia and Manitoba.
HR teams should hold employers to the promise of better jobs for the people who stay. Research covered by HRD found that working alongside robots can erode employees' sense of autonomy and meaning, so the redesigned human role needs as much planning as the robot.
Read next: AI to create more jobs than it eliminates by 2030: Gartner
How far off is it?
Further than the videos from the robot Olympics might suggest (although if you watch the outtakes, you can see the behind the scenes footage that down’t make good robot worker advertising. In a June update on its autonomous bagging robot, Telexistence posted the footage at normal speed and at eight times speed, and said plainly that the robot is not yet fast enough for real use. Its movements are still jerky. Shift an item a few centimeters and accuracy can drop. Sano described the demo as a starting point.
Gartner expects fewer than 20 companies to put humanoid robots into production in supply chain or manufacturing through 2028.
Fast facts
- More than 300 robots Telexistence says it has deployed with Japanese convenience-store chains
- 45% success for the best DreamZero model on a simulated block-picking task
- 91 of 183 recorded human demonstrations rejected as unusable
- 2029 is Seven-Eleven Japan's target for its Astra humanoid
- More than 11 million: Japan's projected worker shortfall by 2040
- 10% projected fall in US cashier jobs, 2024 to 2034
So, is this the future? For some convenience-store counters in Japan, quite possibly, if Seven-Eleven keeps to its 2029 timeline. For most US employers, the nearer change is the job this trial exposed: the person who shows a robot how to do the work.
Read next: Workers who skip AI tools are paying for it with their jobs
