A March recruitment listing reveals that Xianheng International is offering million‑yuan annual packages to attract VLA large‑model algorithm experts, a clear signal that the company is prioritising the integration of multimodal AI into embodied platforms. The advertised role covers full‑chain development of VLA models and their deployment on real machines, notably robots, underscoring a shift from paper research and cloud prototypes toward on‑device, mission‑oriented systems.
In investor briefings earlier this year the company said it is pursuing “AI large models + industry scenario applications,” aiming to commercialise artificial intelligence across energy, transport and emergency services. Xianheng has articulated an explicit product roadmap: from 2025 it plans to develop a “Centaur” platform — a four‑legged robot combined with a robotic arm and a dexterous hand — and from 2026 to begin application development for humanoid robots, while also promoting air‑ground cooperation between drones and ground robots.
The vacancy and the roadmap matter because they map onto two converging trends in AI and robotics. First, large multimodal models — which fuse vision, language and other sensory inputs — are being adapted from static inference tasks to real‑time control and decision‑making, a technically demanding transition known as sim‑to‑real. Second, Chinese industry policy and private capital are funneling resources into embodied intelligence, where software advances only yield returns when paired with robust hardware, sensors and safety engineering.
Offering premium compensation for VLA specialists highlights a growing talent squeeze. Building reliable, deployable robotic systems requires expertise spanning machine learning, control theory, perception, and systems engineering; such cross‑disciplinary staff remain scarce worldwide. High salaries will accelerate head‑hunting by established tech groups and startups alike, raising costs and intensifying competition for engineers who can bridge models and actuators.
Practical hurdles remain substantial. Deploying large models on robots requires solving latency, power and robustness constraints; ensuring safe human‑robot interaction in crowded or high‑risk environments demands rigorous testing and regulatory approval; and commercial viability depends on reducing hardware and data‑collection costs — problems Xianheng and its peers must address if they want fielded products beyond demos.
Strategically, this recruitment and the stated product timeline illustrate how Chinese firms are betting that the next phase of AI value capture will be won by companies that master embodied deployment. Success would accelerate automation in logistics, inspection, disaster response and energy sector operations, but it would also widen the dual‑use debate as more capable physical agents enter civilian and potentially military domains.
For international observers, the story is a reminder that progress in generative and multimodal AI is no longer confined to cloud‑based chatbots and image synthesis. Companies that pair large models with robust robotic platforms can change who delivers services, how critical infrastructure is monitored, and which firms set the standards for safety and interoperability. The race will be decided as much in factory floors and testing fields as in algorithmic research papers.