--- title: "A Robot Tidies Up the Whole Room on Its Own" date: 2026-11-04 category: Humanoids & Robotics site: NeuroAI canonical: https://neuroai.site/a/na-embodied-longtask-16-1104 language: en --- # A Robot Tidies Up the Whole Room on Its Own > Say "tidy up" to a messy living room and the robot sets itself in motion: it recognizes the cola, bowls, toys, and trash, plans an order, moves its body, and puts everything back where it belongs. A Say "tidy up" to a messy living room and the robot sets itself in motion: it recognizes the cola, bowls, toys, and trash, plans an order, moves its body, and puts everything back where it belongs. A scene like this is no longer just a sci-fi movie trope. At a recent exhibition, homegrown Chinese robots demonstrated this small chore of tidying a room with real conviction. Over the past year, humanoid robots have repeatedly stolen the spotlight: dancing on the Spring Festival Gala, guiding visitors in malls, interacting at exhibitions—no shortage of attention, no lack of applause. Yet after all the excitement, we've yet to see them genuinely take on real work. So "can it reliably finish one long job?" has become the new yardstick for measuring value. Some in the industry even wonder whether humanoid robots have reached their limit. ## Autonomous Reasoning in Long Tasks Qianxun Intelligence's Moz1 is a representative of this kind of demo. Given a single natural-language instruction, it can break the task down on its own: put the cola in the fridge, the bowls in the dishwasher, the toys back in place. More importantly, when staff moved the laundry basket on the spot, it could change its route and re-lock onto the target without being given a new instruction. Behind this performance is a general robot agent architecture. A long-horizon "brain" keeps the overall goal and progress in mind, a perception system builds a spatial map and plans routes, and a foundation model generates actions from language, vision, and its own state. Only with multiple stages continuously closing the loop can a robot move step by step from following a script to figuring things out for itself. A little over a month ago, Moz1 publicly demonstrated living-room tidying at the World Artificial Intelligence Conference; at the World Robot Conference, it went up against itself again in a kind of long-distance contest. In an office-scenario event at another competition, Moz1 relied on the same brain, stayed off prewritten scripts the whole way, and took the top spot in the general service role. From living room to office, the setting changed, but the underlying reasoning had to bear the same complexity. On the same stage, the Moz2 commercial service robot also upgraded its voice interaction and collected 999 "love logs" on site, used to polish a friendlier manner for service scenarios. Qianxun has ordered its own path: industry first, then commercial service, then the home, letting the two types of robots validate complex manipulation and public interaction respectively. Stepping away from prewritten scripts and remote control, and completing tasks through the model's own reasoning, is the required course for humanoid robots to enter real life. Qianxun's performance at the competition is seen within the industry as a vivid sample of this direction. ## An Efficiency Jump in Thirty Days The speed of progress is worth noting too. According to Qianxun Intelligence's disclosures, over about thirty days the team kept tuning task understanding, path planning, action execution, and exception handling, and the overall efficiency of Moz1's long-horizon "living-room tidying" task improved by more than 15% compared with before. On specific subtasks, success rates for things like putting cola in the fridge and bowls in the dishwasher rose to above 99%, and the time needed to confirm a navigation target was nearly halved. The numbers look good, but the meaning behind them matters more: it isn't just that individual motions got faster—the handoffs between understanding, navigating, manipulating, and correcting became smoother. Of course, these figures come from the company's own demos and disclosures. What they prove is that it can do better in a prepared environment, not that it's steady in any environment. For readers, telling those two apart matters more than remembering that handsome percentage. ## Tidying a Room in Fifteen Minutes Another young domestic company, Moqi Intelligence, brought the same test question to the exhibition floor. Just a year old, it unveiled its self-developed embodied-intelligence model architecture MoRA and introduced the MORPHI KINO wheeled robot built for long-horizon household tasks. The company says that in a simulated home environment it autonomously completed a housework routine lasting about fifteen minutes, spanning multiple areas and many steps. On camera, it first went to the coffee table, recognized clutter such as crumpled tissues and mineral-water bottles and dropped them in the trash, then gathered scattered items onto a tray; next it moved to the fridge, checked drink stock, fetched mineral water to restock when supplies were low, and closed the door; finally it came to the laundry area, took out the dried clothes, started a drying cycle, and folded and stacked them. No one directed it at any point. Moqi takes an "agent-native" route, pushing goals, memory, and progress down into the execution engine so the robot advances the task itself through continuous motion. Fifteen minutes strings together three everyday chores: cleaning, restocking, and folding. For ordinary people these are unremarkable, but for a robot they're a combined exam in perception, planning, and continuous execution. Long tasks are hard because they link perception, decision-making, and execution into a chain that can't break; if any one link jams, the whole effort is wasted. ## Does It Still Work in a Different Room? At the second World Humanoid Robot Games, scenario events grew to more than twenty, putting more weight on long workflows, multiple tasks, and full autonomy. Behind this is a clear industry shift: humanoid robots are slowly moving from performers under the spotlight toward labor in factories and homes, and events at exhibitions and competitions are edging toward real scenarios too. The sector runs hot, yet genuinely stable deployment is still in its infancy. Even so, one consensus keeps coming up in the industry: completing a task that was prepared in advance isn't the same as still succeeding in a different room with a different cup. Generalization has yet to be fully proven. Qianxun's strong results in two very different settings with the same "brain" are a meaningful exploration of generalization; but these are still carefully designed test questions, a long way from working out of the box. Take a simple example: if the cup used in training is white, switching to a black one with a slanted rim may leave the robot unable to recognize it or get a firm grip. The variables in the real world are endless, while the certainty inside a lab is limited. Success at an exhibition is a carefully built stage; real life never rehearses in advance. An extra sock on the floor, a cabinet door left ajar, a cup in a shape never seen before—any of these can jam a process that had been running smoothly. Whether the capability is truly steady only shows in an environment no one has tidied up for it. The demos from Qianxun and Moqi are more like making the test public and laying the shortcomings bare. That they can work at an exhibition is a step worth applauding; but being equally dependable in someone else's home, in an unrehearsed environment, is the real exam of their value—and the last hurdle in this long-task problem. For ordinary people, rather than asking "when will a robot come into my home?", it's better to first watch whether it can reliably finish those fifteen minutes in more unfamiliar settings. There's still no fixed timetable for when this kind of ability will enter daily life. The answer is being written in every replanned route and every unfamiliar object picked up. **Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.** --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-embodied-longtask-16-1104 Free to quote with attribution and a link to the original.