A Step Toward Dependable Robot Helpers, Starting at a Mall Juice Counter

Think about the people who might one day work beside a robot: someone behind a kitchen counter, someone on a factory line. For them, a cleverly trained machine is no help if it fails one time in twenty. Most learned robot skills look great in a video and still fail often enough that nobody would trust them with real work.

A paper published in Science Robotics on July 22, 2026, reports a robot that kept its footing. A juicing robot running a learning method called RL-100 served customers in a shopping mall continuously for about seven hours without a failure, and nobody retrained it for the mall first. The method comes from Shanghai Jiao Tong University, Tsinghua University, the Shanghai Qi Zhi Institute and other institutions, and was first shared as a preprint on arXiv.

How a robot learns a chore this way

RL-100 starts from human demonstrations, so the robot begins with a reasonable idea of the task. Then it improves through reinforcement learning: it practices, gets scored on success, and adjusts. Both stages run under a single training objective inside a diffusion policy, the family of generative models used to make images, here producing robot motions from what the cameras see. Diffusion models are usually slow, so the team distilled the process into a one-step controller. In the preprint, the robot arms receive new motion commands 30 times a second, fast enough for real-time control.

In the lab, a clean sheet with a frame around it

The robots took on eight real tasks, among them pushing, bowling, pouring, cloth folding, unscrewing, multistage juicing and long-horizon box folding. Liquids, soft cloth and twisting motions are exactly where learned skills tend to stumble. Under a predefined protocol the authors set, robots using the method succeeded in all 1,000 of 1,000 evaluated episodes, including 250 consecutive successes on one task.

The team also checked what happens when the world refuses to cooperate. Without retraining, a single policy succeeded about 90 percent of the time when the environment or the physics changed. After a short period of extra training on substantial task variations, it reached 86.7 percent. When people deliberately disturbed it, it held up about 96 percent of the time.

Next to the people who taught it

The journal paper reports that the robots matched or beat expert human teleoperators, people steering a robot remotely, on time to completion. The preprint adds two head-to-head contests. In tabletop bowling, with the same number of attempts on each side, the robot scored 25 successes against 14 for human players. In a block-pushing task with the same time budget, it completed 20 successful trials, compared with 17 for the expert who had teleoperated the robot to collect its training demonstrations and 13 for a novice operator. For the person whose demonstrations got the whole thing started, that is a funny sort of compliment.

What a shop owner would still want to know

The 1,000 of 1,000 figure belongs to the evaluated trials on those eight tasks, under the authors’ own protocol, not to robots in general. And about seven hours is one long shift on one task. Anyone thinking of handing a robot real work would want months of it, across many sites and tasks, with someone other than the builders keeping score. The sources do not name a commercial deployment.

The authors write that their results suggest a potential path to deployable robot learning by starting from human priors and aligning training with human-grounded metrics. Picture the owner of a small juice stand deciding whether to let a robot take the morning rush, every morning, for a season. That is the test still to come. This paper is a credible first shift.