Qingcang Robotics’ 0.8B VLA model reportedly deployed at L’Oréal’s Suzhou plant; on-site rollout took 30 days

TMTPost reported on June 18, 2026, that Qingcang Robotics’ technology had been deployed at L’Oréal’s Suzhou plant. According to the outlet, as of June 17, on-site data collection, further adaptation of the model and deployment and adjustments using the company’s VLA model, EVO-1, had been completed in 30 days. The model has 0.8B parameters and is said to perform inference using the robot’s onboard computing resources.

TMTPost reported that a 0.8B-parameter model performs inference on the robot and is further adapted using on-site data. If the report is accurate, this points to an option for operating with computing resources at the site rather than relying on large external infrastructure. However, the report provided no comparative figures for deployment costs or operating performance.

The point of interest is less the robot’s task performance than the reported deployment process, which took 30 days from model adaptation through on-site adjustments. TMTPost compared this with a three- to six-month deployment period for conventional industrial robots and attributed the shorter timeline to EVO-1’s small-sample adaptation and on-device inference. However, the excerpt does not specify the start and end points of the timeline or the conditions for the comparison, so the 30-day figure alone cannot establish whether the approach can be replicated at other factories. The reported technical design uses a VLA model that processes vision, language and action to generate action sequences from inputs. The outlet describes this as differing from approaches that rely on programs or dedicated fixtures for each individual process. This is linked to the claim that one robot model and one model can handle multiple tasks. At the same time, the excerpt provides no figures on operating hours, throughput or stoppage frequency for the outlet’s cited capabilities, such as anomaly detection and handling multiple product types. The breadth of tasks and the reliability of sustained operation therefore need to be assessed separately. A related article from this newspaper, Skild AI unveils S1, which performs long-duration tasks of up to 10 minutes from video demos (Aug. 18, 2026), covered the announcement of a foundation model that generates actions for unseen tasks based on video demos. This case, by contrast, is a reported deployment at a factory, so the two can be compared as information about different stages: task execution in a demo and deployment in a production setting. However, differences in model size and task conditions mean they do not provide a basis for direct performance comparisons. The industry implication is that if lightweight models are to handle on-site adaptation, deployment costs and responses to process changes need to be considered as part of an end-to-end operation that includes data collection, model updates and on-site adjustments—not just the robot itself. The report does not say how often the model will be updated after deployment, how much data each task requires or how much human intervention is needed. Points to verify next include the scope of the 30-day deployment period, performance across multiple processes, throughput and stoppages, safety measures for abnormal conditions, and the conditions needed to reproduce small-sample adaptation. Only then will it be possible to determine whether rapid deployment is limited to a particular site or can be applied across processes.