The ROS-Industrial Consortium Asia-Pacific recently shared something that doesn't happen often in enterprise robotics: a clear admission that their existing workflow system wasn't working, and a detailed account of what they replaced it with.
Teams managing multi-agent robotics systems - think factory floors with dozens of robots coordinating tasks - have traditionally relied on Behavior Trees to orchestrate complex workflows. On paper, Behavior Trees make sense. They provide a visual, hierarchical way to define robot decision-making. In practice, at industrial scale, they become unwieldy.
The Problem With Trees at Scale
Here's what the consortium found: when you're managing multiple robots working on interdependent tasks, Behavior Trees start to fracture. Each robot runs its own tree, but coordination between trees requires constant message-passing and state synchronisation. Add more robots, and the complexity compounds. Debugging becomes archaeology - tracing through layers of nested nodes to find where coordination failed.
For a single robot picking and placing objects, a Behavior Tree works brilliantly. For ten robots working in sequence across a production line, with handoffs and shared resources, the tree structure fights you. You're trying to model parallel, asynchronous workflows in a system designed for hierarchical decision-making.
Enter Crossflow: Workflows, Not Trees
The consortium migrated to Crossflow Executor, a relatively new approach within the ROS ecosystem that treats multi-agent coordination as a workflow problem, not a decision tree problem. Instead of each robot managing its own nested logic, Crossflow orchestrates tasks at a higher level - defining what needs to happen, in what order, with what dependencies, and letting the executor handle distribution and state management.
Think of it like moving from microservices where each service manages its own state, to an orchestration layer like Kubernetes that manages workloads centrally. You define the workflow once, and the executor handles execution across multiple agents.
The presentation covered real industrial use cases - not toy examples. Assembly lines. Warehouse logistics. Scenarios where timing, handoffs, and error recovery matter. The kind of workflows where a single missed handoff can halt an entire production line.
What This Means for Builders
If you're building multi-agent robotics systems, this migration pattern is worth studying. The shift from Behavior Trees to Crossflow isn't just a technical preference - it's a recognition that workflow orchestration and individual robot decision-making are different problems requiring different tools.
Behavior Trees excel at reactive, hierarchical control for single agents. Crossflow excels at coordinating parallel, interdependent tasks across multiple agents. Most industrial robotics deployments need both - local intelligence and global coordination.
The consortium's documentation is practical and detailed. They're not selling a vendor solution - they're sharing what worked in production after the previous approach hit scaling limits. For anyone deploying robotics at scale, that kind of honest, use-case-driven engineering insight is rare and valuable.
Industrial robotics is finally moving beyond isolated, pre-programmed robots toward genuinely coordinated multi-agent systems. The tooling is catching up to the ambition. Crossflow is one piece of that puzzle - not the only answer, but a pragmatic one that's solving real workflow complexity in production environments.