The competitive focus within the embodied intelligence sector is shifting from demonstrations of mobility to validation of real-world operational capabilities.

On August 19, the 2026 World Robot Conference opened in Beijing. Unlike previous years which showcased running, dancing, and similar athletic feats, this year's event sees a growing number of companies bringing warehousing and manufacturing workflows directly into the exhibition halls. GalaxyX showcased robots completing tasks such as product picking, sorting, bagging, sealing, and screw fastening.

Gao Jiyang, CEO of GalaxyX, stated that on the day before the conference, the live system built in collaboration with JD.com for a front-warehouse operation successfully fulfilled over 100 orders with zero manual intervention throughout the entire process.

Gao also revealed that this year, the company has begun delivering thousand-unit-scale, model-driven production units. He specifically distinguished this metric from standard robot sales, clarifying that the statistics only encompass "model-driven productive industry units."

Gao emphasized that GalaxyX currently evaluates its performance based on two key indicators: model capability and the progress of commercial deployment driven by that model capability. He noted, "Customers are not buying a robot machine; they are purchasing the production work the robot can complete. What clients pay for is labor output."

However, transitioning from completing tasks on an exhibition floor to establishing stable production capacity at a customer's site still requires significant engineering and commercialization efforts. Whether robots can operate continuously, reduce the need for human intervention, and keep unit operational costs within acceptable limits for clients remains a key challenge for GalaxyX and the industry at large.

Entering the Workflow

At this year's World Robot Conference, demonstrations of depalletizing, palletizing, retail picking, material handling, and assembly were prominently featured.

Gao noted that starting from industry events like last year's World Robot Conference, the sector's focus has shifted toward intelligent and model-driven capabilities, with attention turning to "whether robots can actually perform real work." This trend has become even more pronounced this year.

Running, jumping, and navigating obstacles primarily test motion control. But once robots enter warehouses and factories, they must understand tasks, manipulate objects, handle anomalies, and integrate with existing production systems. Accordingly, evaluation criteria have evolved to include task success rates, continuous operation time, and unit operational costs.

The front-warehouse operation demonstrated by GalaxyX consists of multiple sequential steps. Facing products of varying sizes, materials, and placements, the robot must perform picking, sorting, placing, and packaging, while using dual-arm coordination to open bags, load items, and seal them. The company claims this live system can handle thousands to tens of thousands of product types.

The assembly station tests a different set of skills. The robot must pick up screws, align them with holes, and operate a screwdriver to complete the fastening process.

Following pre-training on foundation models, GalaxyX incorporates supervised fine-tuning, imitation learning, and reinforcement learning to improve operational precision and task success rates. When anomalies occur during robot operations, human staff can take over, and the resulting data is fed back into the model iteration pipeline.

On the model development front, GalaxyX is simultaneously advancing vision-language-action models and world models.

Gao explained that both model types require converting inputs into tokens, encoding them through multi-layer Transformer neural networks, and pre-training using self-supervised methods. The primary difference lies in the training approach, with world models relying mainly on video prediction. He predicts these two technical paths will eventually converge.

At this year's conference, GalaxyX also unveiled its wheeled-armed robot, Nexo. According to company specifications, Nexo features 30 degrees of freedom, a maximum payload of 10 kg per arm and 20 kg for both arms combined, and an 8-hour battery life, primarily targeting structured environments such as e-commerce retail, industrial manufacturing, and express logistics.

GalaxyX is also developing the bipedal robot Kengo and the developer-focused product Lemo.

Under Gao's "1+3+N" framework, wheeled-armed robots focus on productive operations, Kengo targets unstructured and complex environments, and Lemo aims to lower the barrier for developers using models and hardware.

This product portfolio reflects a practical reality: different scenarios require different robot form factors.

Warehouses and factories prioritize payload capacity, endurance, and continuous operation, while environments with stairs and ramps demand greater mobility capabilities. For customers, the robot's appearance is not the primary concern—whether it can reliably complete the work is the foundation for deployment.

From isolated actions to complete workflows, GalaxyX is pushing robots into real production environments. However, whether thousand-unit deliveries translate into stable production capacity will depend on continuous operation time, task success rates, and human intervention rates at customer sites.

The Challenge of Cost-Effectiveness

Gao divides the commercialization of embodied intelligence into several phases.

Before 2026, model capabilities were relatively limited, and market transactions primarily involved whole-machine sales. From the second half of this year through 2028, the industry is expected to gradually see subscription-based models, where clients pay for industry-specific solutions for particular production scenarios, with the machine serving merely as the physical carrier of the solution.

As intelligent capabilities expand from single scenarios to multi-scenario cross-domain applications, GalaxyX anticipates that the business model for embodied intelligence may further evolve to center on "physical world token sales," with the industry's value gradually shifting from hardware to intelligent capabilities.

This remains GalaxyX's projection for future business models, contingent on model capabilities and customer acceptance.

When customers calculate whether robots are economically viable, comparing only the machine's price against labor wages is insufficient. Equipment depreciation, energy consumption, maintenance, post-training, system integration, and on-site support must all be factored into costs. If robots require constant monitoring or frequent human intervention, the associated labor costs will also affect the overall return on investment.

Gao noted that current robot operational efficiency is approximately 70% to 80% of human capability, and adapting to a new task requires roughly 10 hours of training.

However, the 70% to 80% efficiency figure serves only as a baseline for comparison and cannot be directly converted into ROI. The complexity of actions and anomaly rates vary significantly across different tasks.

The cost of learning new tasks also impacts deployment speed. GalaxyX currently requires about 10 hours of post-training to adapt to a new task, with the goal of compressing this to 1 hour, or even requiring only a few samples. Whether this training threshold can be lowered will determine if a model can be replicated cost-effectively across more clients and workstations.

Regarding data strategy, GalaxyX insists on prioritizing real-world data. Gao stated that the company relies almost exclusively on real data, rarely using simulation data, and therefore encounters fewer issues with sim-to-real transfer.

On costs, he emphasized the need to combine three components: data, computing power, and R&D engineer manpower. According to GalaxyX's assessment, data costs are relatively controllable, "computing power is the most expensive, and talent is the scarcest resource."

GalaxyX also plans to combine crowdsourced data collection with its internal professional team. In June, the company launched its "Million Hours of Real Data Initiative." Gao did not disclose current data volumes, only noting that one million hours represents the scale threshold the company believes is necessary to support an industrially valuable foundation model, and that existing volumes remain insufficient.

Gao believes the technological inflection point for embodied intelligence will arrive, but public perception will not shift overnight. Robots will first enter factories and warehouses, and industry changes are more likely to manifest as sustained penetration with workstations gradually opening up one by one.

For GalaxyX, the next phase of validation extends beyond whether robots can complete tasks, to whether these capabilities can be reliably replicated across different customer sites, ultimately reducing unit operational costs to levels clients find acceptable.

The evaluation standard for embodied intelligence commercialization is also shifting from "can it do the work" to "is it more economical than existing methods." Machine pricing, operational efficiency, training, and maintenance costs collectively determine unit operational costs and will influence the pace at which robots are deployed at scale.