We Connected Digital Agents to a Physical Robot at WRC 2026 — Why Robots Need Two Brains

At the World Robot Conference 2026, Mininglamp and HIKROBOT demonstrated a robot that assisted with the award ceremony, showcasing a joint booth where three scenarios—restaurant cleaning, warehouse logistics, and patrol—were built around Mininglamp founder Minghui Wu's keynote argument that robots need two brains to enter commercial production. The second brain, as explained by Mininglamp, is an organizational brain that handles multi-machine coordination, cross-system dispatch, and integration with existing IT systems, complementing the VLA model's perception and control capabilities. Mininglamp's open-source platform Octo, which enables digital agents to collaborate, was extended to orchestrate physical robots in this demonstration.

At this year's World Robot Conference WRC in August, a robot stood beside the podium on the closing day, helping staff carry out the award ceremony. It was a joint creation from Mininglamp and HIKROBOT's shared booth, and one of the few images from the event that stuck with both trade visitors and camera lenses. The robot's significance wasn't in the ceremony itself. It appeared on that stage because of a keynote given earlier, where Mininglamp founder Minghui Wu laid out a judgment at the WRC main forum: for robots to actually enter commercial production systems, they need two brains. The three scenarios shown at the booth — restaurant cleaning, warehouse logistics, and patrol — were all built around that judgment. This piece isn't really about the keynote, or the robot. It's about the judgment itself, and what "the second brain" actually means when we say a robot needs one. The hottest battlefield in embodied AI for the past two years has been the pipeline from visual input to physical action. The names everyone in the field recognizes sit on this line: OpenVLA, the π0 series, VLA-JEPA. What they're all doing, at core, is letting a robot see, understand, and act, the way a person would. Progress here has been fast — new data scales, new architectures, new training strategies every few months. But put these models into an actual restaurant, an actual warehouse, an actual patrol route, and a different layer of problems shows up immediately: None of these questions belong to any single VLA model's job description. They live above the model, not inside it. That's what we mean by the second brain — the organizational brain. The organizational brain handles three things: coordinating multiple machines, dispatching across systems, and connecting with a company's existing IT stack. Multi-machine coordination means multiple robots know what each other is doing, avoid conflicts, and share context. Cross-system dispatch means digital agents and physical robots can hand tasks to each other and exchange state. Connecting with existing IT means a robot stops being a standalone showpiece and becomes one link in a scheduling system, an inventory system, a ticketing system. Making the VLA model bigger and stronger doesn't solve any of these. They're orchestration problems, not perception or control problems. This isn't a new judgment for us. Mininglamp put forward a framework back in 2018 called HAO — H for Human, A for a broad notion of artificial intelligence covering both digital agents and physical robots, O for Organizational Intelligence. The core idea was to connect people, agents, and robots onto one network, generating organization-level productivity through division of labor rather than expecting any single entity to be fully capable on its own. Octo, Mininglamp's open-source human-agent collaboration platform, is the first step of that idea being built in the digital world — letting digital agents collaborate with each other, share context, and accumulate experience. The three scenarios shown at the joint WRC booth were the first public demonstration of that orchestration capability extending from the digital world into the physical one. Line up the technical evolution of agents against the technical evolution of robots, and something interesting shows up: they're following the same road. On tool use: the digital world moved from rule engines, to expert systems, to RPA, to workflows, and eventually to general-purpose agents. The embodied world moved from PLC programming, to behavior trees, to end-to-end VLA models today. Both lines swing between "hard-coded" and "fully autonomous." The digital world eventually landed on something in between — structured skills, with Anthropic's Claude Skills as a recent example. The embodied world has started producing something similar, structured "atomic skill libraries" as a complement to pure end-to-end approaches. The memory line is even more clearly parallel. Digital agents use working memory, episodic memory, semantic memory, and procedural memory — the CoALA four-layer framework. Embodied robots need something equivalent: spatial memory, task trajectory memory, semantic memory, skill memory. It's the same memory architecture, sitting on two different bodies. There's one place that still needs bridging: "ontology" in embodied contexts usually means physical structure, while ontology in semantic memory means an abstraction of concepts and relationships. Eventually these two need to share one semantic layer for a robot to actually understand what a human or an agent is saying. This is also why we think a team that has spent years building digital agent orchestration isn't crossing into a new field by working on embodied intelligence — it's a natural extension of the same work. For the organizational brain to actually function, it needs a communication protocol connecting robots, digital agents, and legacy systems. The ideal shape of that protocol looks a lot like email — open, simple, not locked to a single vendor. Every company should be able to run its own orchestration platform, and those platforms should still be able to talk to each other. The digital world has made progress on this front over the past year or two: MCP addresses the protocol between models and tools; A2A addresses the protocol between agents. Once robots join the picture, the protocol has to cover more types of participants, and the requirements on latency and security get stricter — but the principle of staying open and vendor-neutral should hold across both worlds. The digital world has broadly gone through five stages: L1, everyone has access to usable model inference; L2, everyone works with agents day-to-day, and legacy tools become callable by agents; L3, humans and agents share structured semantic memory; L4, a complex collaboration network forms across the workforce, continuously improved through evaluation and benchmarking; L5, a company accumulates its own data and capability, decoupled from any specific base model. Whether this path repeats itself in the embodied world isn't something we have an answer for yet. Embodied AI as a field is still young, and the infrastructure layer is still being figured out. We're putting this out as a hypothesis worth discussing, not a conclusion we've already validated. Buying an AI tool doesn't automatically mean a business process starts running. Someone usually has to be on-site, fitting the model's capabilities and skills into a company's actual workflow. The digital world already has a name for this role: the Forward Deployed Engineer FDE . Palantir built its early success on this role in complex, real-world deployments, and it's since become an industry-recognized position. Robots entering commercial settings need the same kind of role — someone who connects motion control capability with the actual operating procedure of a specific scene, so the robot becomes a working part of the process rather than a demo on a show floor. This layer of capability comes from engineering practice, not from the model itself. The "organizational brain" discussed at WRC is, in engineering terms, exactly what Octo has been building as its orchestration layer. Up to now, Octo has mainly connected digital-world agents — letting different agents divide up complex tasks, share context, and accumulate experience. Looking ahead, there's no reason for this orchestration capability to stay confined to the digital world. Robots and embodied devices are, in principle, participants that can be brought into the same collaboration network; the difference lies in how they connect and what constraints apply — physical actions are often irreversible, and latency requirements are stricter, both of which the orchestration layer needs to account for. We currently see this WRC collaboration as a validation of direction, not a finished answer. Restaurant cleaning, warehouse logistics, and patrol are concrete business collaborations between Mininglamp and HIKROBOT; how far the organizational brain layer can actually go still needs more time and more real-world scenarios to work out. That's also why we keep Octo open source, and why we're putting this judgment out for discussion — a genuinely open orchestration protocol should be tested by enough scenarios and enough developers, not settled behind closed doors. If you're working on multi-agent orchestration, embodied AI deployment, or robot-agent communication protocols, feel free to check out the Octo repo and open an issue: