At this year's World Robot Conference in Beijing, a robot walked up to the podium on closing day and helped run the award ceremony. It came from a joint booth between Mininglamp and HIKROBOT.
The robot mattered less than the reason it was there. Four days earlier, at the main forum, Mininglamp's founder laid out a claim: for robots to actually enter commercial production, they need two brains. The three scenarios on the booth — restaurant cleaning, warehouse logistics, patrol — were all built around that claim. This post is about the claim.
The First Brain: Vision In, Action Out
The hottest battlefield in embodied AI has been the pipeline from visual input to physical action. OpenVLA, the π0 series, VLA-JEPA — all of them are trying to let a robot see, understand, and act the way a person would. Progress here is fast.
Push those models into an actual restaurant or warehouse and a different layer of problems shows up:
One robot clears tables, another collects trays. How do they coordinate?
The front-of-house system fires a new order. How does the kitchen robot know to start prep?
A patrol robot flags an anomaly. How does that event reach security and the person on duty?
None of these live inside a VLA model. They live above it. That upper layer is the second brain — the organizational brain.
The Second Brain: Orchestration Above the Model
The organizational brain does three things: coordinate multiple machines, dispatch across systems, connect to existing IT. Making the VLA model bigger doesn't solve any of these — they're orchestration problems, not perception problems.
This isn't a new idea for us. In 2018 Mininglamp put out a framework called HAO — Human, a broad notion of AI (digital agents and physical robots both count), and Organizational Intelligence. The point was to put people, agents, and robots on one network and let division of labor produce organization-level output.
Octo, Mininglamp's open-source human-agent collaboration platform, is the first step of that idea in the digital world — agents dividing tasks, sharing context, accumulating experience. The WRC scenarios were the first public demo of that orchestration extending into the physical world.
Same Road, Two Bodies
Line up the technical evolution of agents against robots and you see the same road.
Tool use: the digital world went rule engines → expert systems → RPA → workflows → general agents. The embodied world went PLC → behavior trees → end-to-end VLA. Both swing between hard-coded and fully autonomous, and both are landing on structured skills in the middle (Claude Skills on the digital side, atomic skill libraries on the robot side).
Memory is even more parallel. Digital agents use working, episodic, semantic, and procedural memory — the CoALA four layers. Embodied robots need spatial memory, task trajectory memory, semantic memory, skill memory. Same architecture, different bodies.
One gap worth naming: "ontology" in embodied contexts usually means physical structure, while ontology in semantic memory is an abstraction of concepts and relationships. Those two eventually need to share one semantic layer, or a robot can't really understand what a human or an agent is saying.
A Protocol Shaped Like Email
For the organizational brain to actually run, robots, agents, and legacy systems need a shared communication protocol.
The shape we'd argue for looks 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 talk to each other.
The digital side has been moving. MCP addresses model-to-tool. A2A addresses agent-to-agent. Once robots join, participant types expand and latency and safety tighten — but the open, vendor-neutral principle should carry across both worlds.
Where We Actually Are
We treat the WRC collaboration as a validation of direction, not a finished answer. How far the organizational-brain layer can actually go still needs more time and more real-world scenes.
Two things worth being honest about. Physical actions are often irreversible, and latency budgets are tighter than anything the digital-agent world has had to deal with — the orchestration layer has to account for both. And the L1-to-L5 organizational path we've watched play out in the digital world (usable model access, daily agent use, shared semantic memory, evaluation networks, base-model-decoupled capability) — whether it repeats in the embodied world is a hypothesis, not a conclusion.
That's also why Octo is open source. A genuinely open orchestration protocol has to 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, the repo is open:
Top comments (1)
The two-brains framing becomes especially important once a physical action can hurt someone. The organizational layer needs more than routing: explicit capability grants, freshness bounds on world state, conflict resolution, and a local stop path that does not depend on the network. An open protocol should carry those constraints with the command, not leave each robot to infer them from prose.