Westlake Robotics Upgrades General Brain WR1: Two-Robot Folding and Three-Room Chores Demo
According to an October 10, 2026 report, Hangzhou-based Westlake Robotics unveiled an upgraded version of its general-purpose brain model WR1, combining a world model with a VLA architecture. In a continuous single-take demonstration, the system carried out a sequence of tasks: picking up laundry, two humanoid robots collaboratively folding trousers, retrieving corn and placing it in an air fryer, tidying up plush toys, and making a bed, spanning the kitchen, living room, and bedroom. The report claims the dexterous-hand autonomous clothes folding is the world's first such demonstration.
What changed
According to an October 10, 2026 report, Hangzhou-based Westlake Robotics, a company spun out of AI and robotics work at Westlake University, presented an upgraded version of its general-purpose brain model WR1. The report states that WR1 uses a general-purpose dual pre-training architecture for a "large brain" and "small brain," combining a world model with a vision-language-action model to output a complete action sequence covering body movement, full-body posture, and dexterous hand manipulation based on visual observations and language instructions, with full-body control executed by its self-developed general motion model GAE. The demonstration had two parts. In the first, the robot took a pair of trousers off a drying rack and draped them over its arm; despite disturbances such as the rack swaying and the positions of two T-shirts on it shifting, it continued to observe and adjust. It then tossed a T-shirt to a robotic arm for autonomous folding, turned, and spread the trousers on a table to fold them collaboratively with another humanoid robot. The report claims that dexterous-hand autonomous clothes folding by a humanoid robot is the world's first such demonstration. In the second part, the robot moved across three areas -- kitchen, living room, and bedroom -- sequentially opening the refrigerator to take out corn, making a U-turn to the air fryer to place the corn inside, crouching to tidy plush toys on the sofa and coffee table, and making a bed. The report claims no manual resets and no segmented instruction issuing were needed throughout. The report also mentions that WR1 introduces long- and short-horizon motion state review and real-time action chunking to reduce pauses caused by action updates, with an adjustable lookahead mechanism to compensate for execution latency. The source mentions the company has completed multiple funding rounds in the hundreds of millions of yuan, but did not disclose model parameter count, training data scale, or inference latency.
What it means for developers
First, the category must be defined: WR1 is a hardware-coupled embodied intelligence system and does not belong to the same product form as the language and multimodal models listed on KimiSeek, such as DeepSeek, Qwen, GLM, and Kimi. Developers cannot directly use it through existing text or image API call paths, especially its full-body control component. For teams focused on robotics stacks and VLA directions, the analyzable value of this announcement lies in the public description of its architectural layering and task-chaining mechanisms: splitting multimodal understanding, world model prediction, and full-body control into two layers -- a general brain and an execution model -- and then using a task management module to organize the instruction flow. This route aligns with the mainstream direction of current VLA research. The report's claims of "matching or even surpassing Figure and Helix solutions" and "the world's first dexterous-hand autonomous clothes folding demonstration" are unilateral source claims lacking third-party reproduction or metric comparisons, and should not be cited as verified conclusions. In addition, such demonstrations typically occur in controlled sets. Whether cross-room continuity can be stably maintained in unstructured real home environments is not addressed by the source, which provides no failure rate, number of attempts, or indication of whether the footage was edited. Therefore, engineering maturity should not be inferred from this. For model-side developers, the more practical takeaway is the relationship between per-step success rate and overall success rate decay in long-horizon tasks, and whether state-review mechanisms can mitigate pauses at action transitions. These two points share a similar problem structure with long-chain agent orchestration and can serve as analogical references, but they are not equivalent to directly reusable technical solutions.
Integration and evaluation
The following is conditional analysis and advice, not facts stated by the source. First, if a team is working on long-chain agent or robotic task orchestration, it is worth comparing the source's description of long- and short-horizon motion state review and real-time action chunking: incorporating the currently executing trajectory into context when generating the next segment of actions. Whether it is worth introducing depends on your task chain length -- when per-step success rate is high but end-to-end success rate decays noticeably with step count, such state-chaining mechanisms offer the greatest benefit. An evaluation criterion could be the end-to-end success rate curve for the same task as the number of steps increases from three to ten, observing whether the decay slope is flattened. Second, if a team is focused on deformable or flexible object manipulation, the source lists continuously changing object shapes, external disturbances, and dual-robot collaboration as the main difficulties, with the corresponding choice being whether to invest in a perception closed loop or rely on preset trajectories. When the manipulated object is a deformable item such as clothing and there are disturbance sources such as a swaying rack or moving objects in the scene, a closed-loop perception strategy is more likely to be effective. An evaluation criterion could be the drop in task success rate relative to a no-disturbance baseline after deliberately introducing disturbances, for example, whether the success rate decline stays within an acceptable range after shaking the rack or changing the object's initial position. Third, if a team considers reproducing a similar layered architecture, it must first clarify whether it has the conditions to uniformly collect both demonstration and autonomous execution data. The source states that GAE handles both demonstration collection and autonomous execution stages, sharing the same control foundation. This premise does not hold for teams that only have inference-side data and no teleoperation data pipeline for the physical body; in that case, the layered approach should not be copied directly. None of the above suggestions imply that the source provided corresponding data.
Limits and open questions
It must be made clear: this version of WR1 is not currently listed on the KimiSeek platform. What is listed on this site are language and multimodal models such as DeepSeek, Qwen, GLM, and Kimi, which belong to a different category from embodied intelligence systems. If the source later mentions any API or commercial availability, that refers to the vendor side rather than this site, and specific billing and access terms should be based on the vendor's official statements. The source was published by the company itself, and the demonstration is a self-reported performance showcase. It lacks independent third-party reproduction, failure rate statistics, number of attempts, and whether the footage was edited. Statements in the report such as "world's first" and "surpassing the Helix solution" are unilateral source claims and should not be cited as verified facts. The source did not disclose WR1's model parameter count, training data scale, inference latency, whether weights are open or an API is provided, nor the hardware cost and deployment conditions used in the demonstration. Therefore, its engineering usability and reproducibility cannot be judged. The reporting date of this article is October 10, 2026. The source did not separately provide specific dates for when the demonstration was filmed or released; relevant timing should be based on the latest statements from Westlake Robotics' official channels.