Panasonic Holdings team places second in humanoid furniture assembly contest; 34 teams compete

A joint team with the company, centered on the Panasonic Connect Group, also placed seventh.

The competition was judged on the completion of multiple steps and the time taken. Panasonic Holdings’ team placed second by combining step-by-step decisions based on task recognition and failure detection with individual actions. The result is an example of evaluating a control system on assembly work spanning multiple steps rather than on a single action, but it does not demonstrate reproducibility outside the competition or applicability in real-world operations.

Placing second among 34 teams reflects a ranking in a specific furniture assembly competition; it does not indicate the overall performance of humanoid robots or deployment in real-world environments. At the same time, because the evaluation covered multiple steps and included completion status and task time, the competition appears to have tested not only individual actions but also control that breaks down work and carries it through to completion. The design used a high-level planner that recognizes task status and failures to select the next subtask, separating decision-making from the lower-level policy responsible for actions. The technical focus, then, was not just the robot’s physical capabilities but its ability to perform assembly as a system that includes judgments about progress through the steps. The team did not rely solely on its own VLA model, which it designed and trained, but also used existing models such as GR00T for different steps. This should be viewed as a configuration that divides roles among models, not as an announcement confirming that recognition, decision-making and action were consolidated into a single model. Handling different actions—including grasping and inserting parts, and rotating and flipping a tabletop—within the same assembly task served as a test of integrating multiple action elements at the step level. However, the published information does not include scores by step, failure rates, the number of retries or the roles assigned to each model, so the contribution of individual technical elements cannot be assessed from the ranking alone. We reported on April 22, 2026, that Panasonic Connect expands Robo Sync, covering the addition of image processing and recognition functions to a robot control platform for manufacturing sites. In addition, our article published Sept. 24, 2026, [Kindai University improves automatic sorting equipment for red sea bream fry using Panasonic Connect’s collaborative robot control service](/news/kinkidaigaku-panasonikku-konekuto-no-kyoudou-robotto-seigyo-20657a) reported on a trial deployment of the collaborative robot control service. Those reports concerned a control service for use at work sites and improvements to specific equipment; the company has not said they use the same product or platform as the AI control in this humanoid competition. Rather than broadly grouping them as the company’s robotics technology, it is important to distinguish control for work sites from research and development aimed at autonomously carrying out multiple steps. The next points to verify are reproducibility in environments outside the competition and with different furniture and parts, the hardware configuration, results for each step and the scope of the training data. The announcement gives the competition ranking and system configuration, but does not specify plans for mass production or commercial availability, or operating conditions in real-world environments. For now, it is appropriate to assess the result as an achievement in integrated control in a competition, separately from the performance and scope of application of a practical system.

The competition involved assembly work across multiple steps, and teams were evaluated on completion at each step and task time.