Qianxun’s Spirit v1.6 Beats “Goldfish Memory” to Clean a Whole Living Room

During the WAIC exhibition, Qianxun Intelligent displayed two contrasting robots: the task‑focused, black‑gray Moz1 and the friendly, cream‑colored Moz2 with a pink scarf. When staff issued the command “Help me tidy the living room,” Moz1 first scanned the scene, then autonomously moved a soda can to the refrigerator, carried dirty dishes to the dishwasher, and proceeded to collect scattered toys.

Mid‑task, a paper ball was tossed onto the table. Moz1 paused, re‑perceived the environment, reordered its plan, picked up the paper, discarded it, and then resumed the toy‑collection sub‑task without restarting from the beginning, illustrating dynamic replanning without full script reset.

The article explains why “cleaning a living room” is far harder for robots than it appears: humans implicitly apply lifelong common‑sense knowledge to decide what is messy, where items belong, and which subtasks to prioritize, whereas robots lack such priors and must cope with non‑structured, ever‑changing environments with occlusions and unseen objects.

Two inter‑related capabilities are highlighted: long‑range execution, which requires maintaining the overall goal, completed steps, and pending subtasks; and scene generalization, which demands adapting to layout changes, new objects, and temporary disturbances while reusing learned skills.

The article labels the loss of context across steps as “goldfish memory.” It notes that if each sub‑step succeeds with 95% probability, a ten‑step chain has only about a 60% chance of flawless completion, emphasizing the need for failure detection, state monitoring, and on‑the‑fly replanning.

Spirit v1.6 is presented as more than a simple action chain. It integrates a Vision‑Language Agent (VLA) that links language goals, visual perception, and executable actions with a world model that predicts post‑action environmental changes. This integration forms a closed feedback loop for continuous state estimation and plan adjustment.

For example, placing a drink into a refrigerator requires the robot to locate the drink, choose a grasp point, verify the fridge door’s state, execute the grasp, and then re‑observe to confirm the drink is secured and the placement succeeded before proceeding.

During pre‑training, Spirit v1.6 introduces abstract “action tokens” that represent high‑level behaviors, reducing the length of action sequences, lowering computational cost, and strengthening the coupling between perception, decision‑making, and execution adjustments.

The system’s data pipeline is described as a multi‑layered pyramid: internet videos provide generic visual and semantic knowledge; wearable devices capture human‑in‑the‑loop interactions; tele‑operation data helps adapt models to the robot’s hardware; and real‑world robot runs generate success/failure logs that are cleaned, annotated, and fed back for model upgrades. This loop ensures that real‑world anomalies—such as missed grasps or slip events—directly inform subsequent training cycles.

Industrial validation is highlighted: Moz1 is already deployed on CATL’s battery‑pack production line for high‑voltage plug insertion, achieving a success rate above 99%. Partnerships with JD.com, Bosch, and Schaeffler extend the technology to retail and other industrial scenarios, where precision, cadence, and safety are paramount.

In contrast, Moz2 targets commercial service spaces like malls, hotels, and offices. Its design features a soft, rounded chassis, 32 degrees of freedom, omni‑wheel base with foldable legs, and a 7‑DOF biomimetic arm with a three‑finger hand, enabling flexible navigation and future object‑delivery tasks. Currently it serves as a platform for human‑robot interaction and hardware validation, with plans to inherit Moz1’s long‑range planning and dynamic adaptation capabilities.

The roadmap is described as “industrial first, then service, then home,” emphasizing that mature, stable performance in strict industrial settings provides the data foundation needed before tackling the open, highly variable environments of public spaces and ultimately domestic households.

In conclusion, the article argues that robotics is moving from isolated single‑step demos toward embodied intelligence capable of sustaining long‑duration, goal‑directed tasks in real, unstructured environments—much like the early flights of the Wright brothers opened the path to modern aviation.