On August 19, the 2026 World Robot Conference (WRC) opened in Beijing. Mininglamp (HKEX: 2718), together with HIKROBOT, showcased embodied intelligence progress in commercial service scenarios — marking Mininglamp's formal entry into the embodied AI track.
Commercial service robots are currently one of the most closely watched deployment directions in the robotics industry, and 2026 is widely seen as the key commercialization window, with cleaning robots being the fastest-growing sub-category. But on the ground, getting hardware and software to work together — and getting a robot to independently complete long-horizon tasks in real, unstructured environments — is still an open problem.
Mininglamp's answer is to give robots an "AI brain" that can observe, reason, and collaborate, so commercial service robots can autonomously complete full task chains in open, unstructured environments.
On the show floor: a closed loop for commercial service scenarios
At this year's WRC, the joint Mininglamp/HIKROBOT booth focused on commercial service scenarios, demonstrating three use cases live: restaurant cleaning, industrial logistics, and smart patrol.
In the restaurant scenario, a wheeled humanoid robot and a wheeled dual-arm robot worked together to autonomously clear tables, collect dishes, and clean floors — a full closed loop from autonomous task planning to end-to-end execution in an unstructured environment. One detail worth noting: after large-scale training, the robots can distinguish objects on a table, telling food waste apart from items like phones, keys, or documents, and only removing the trash without disturbing anything a customer left behind.
In the industrial logistics and smart patrol scenarios, robots demonstrated a full flexible piece-picking pipeline — high-shelf retrieval, flat-surface transport, case handling, and fine-grained sorting — while the patrol robot showed autonomous patrolling, real-time perception, and anomaly detection in a complex environment.
On the division of labor: Mininglamp focused on VLM/VLA (vision-language models / vision-language-action models), building the core intelligence stack — multimodal perception, task reasoning and planning, and multi-agent coordination — to address the pain points of commercial service robots operating in unstructured settings. Mininglamp's Mingsheng Pinzhi unit has spent years in offline restaurant operations, accumulating data and domain know-how. HIKROBOT brought deep expertise in mobile robot hardware, motion control, multi-sensor fusion, systems software, integration, and mass production engineering. Together, the two sides are exploring how AI-driven commercial service robots can scale.
Why Mininglamp is getting into embodied AI: the organizational intelligence thread
This year's WRC theme was "human-machine symbiosis, supply-demand convergence" — and how humans, digital agents, and other agents collaborate efficiently is something Mininglamp has been working on for a long time.
Back in 2018, Mininglamp proposed the HAO framework (Human + AI + Organization): H stands for Human; A stands for Artificial Intelligence, covering both digital-world agents and physical-world robots; O stands for Organizational Intelligence. Mininglamp's open-source human-AI collaboration platform, Octo, is how HAO intelligence is being put into practice — bringing humans, digital agents, and robots onto the same collaborative network, sharing context, tasks, and preferences.
In Mininglamp's view, intelligent labor inside an organization is now showing up in two forms at once: agents in the digital world, and robots in the physical world. Though they belong to different worlds, the underlying technical paths are structurally similar — multimodal perception, task reasoning and planning, and action execution are, at their core, the same capability stack applied to two different "bodies." Mininglamp previously taught software-world agents to see, think, and act; this time, the same capability is extending to physical-world robots, letting them participate in collaboration as members of the organization.
The second half of robotics is about the "brain"
Mininglamp founder, CEO and CTO Wu Minghui gave a keynote at the WRC main forum, titled "Agent + Embodiment: How Application Demand Drives the Evolution of Robot Intelligence," laying out Mininglamp's thesis on the embodied AI track.
"The key to the second half of robotics is the brain," Wu said. He described two dimensions of "brain": the first is the traditional robot brain — reasoning, task decision-making, and motion planning capability, powered by models like VLM and VLA. The second is the organization's brain. Once a robot is actually deployed into a production system, into offline service industries and commercial scenarios, how it gets embedded into the broader organizational system — eventually forming an organization-level brain — matters just as much. No single robot's own model can solve that by itself.
That's exactly where Octo, Mininglamp's open-source human-AI collaboration platform, comes in. As Wu put it, what Octo is ultimately building toward is "an organization-level human-machine collaborative brain, not just the individual brain of each robot." He also spoke about future connectivity: "We believe Octo will eventually connect not just agents in the digital world, but agents in the physical world too — embodied devices, self-driving cars. Robots to robots, robots to agents, legacy IT systems to next-generation AI systems — in the future, all of this might be connected through open protocols."
From the digital world to the physical world
From unmanned restaurant cleaning to flexible warehouse picking, embodied AI is moving fast from technical validation to industrial deployment. For Mininglamp, this isn't just about entering a new track — it's a validation of a transferable capability. The multimodal understanding, reasoning, and multi-agent coordination refined in the digital world are now being reused and extended into physical-world robots; in turn, the data and experience robots accumulate in real scenarios will feed back into model iteration, forming a positive flywheel of scenario, data, and capability.
That's how "organizational intelligence" moves from blueprint to reality: get an agent's perception, decision-making, and collaboration working in the digital world first, then extend the same intelligence to every robot in the physical world. Octo is open source on GitHub — if you're interested in how humans and AI agents (and eventually physical-world agents) can collaborate, take a look at the code, and a star is always appreciated: https://github.com/Mininglamp-OSS/octo-server
Top comments (0)