Getting one robot to behave intelligently is hard enough. Getting a mixed team of them to coordinate, adapt on the fly and pursue a shared goal is a different order of problem entirely — and it’s exactly what researchers at the Johns Hopkins Applied Physics Laboratory (APL) have been tackling with a new push into agentic AI.
On June 17, 2026, APL presented its work in a free webinar that laid out both the ambition and the messiness of the field. The pitch, in short: take the reasoning power of large language models and hand it to groups of heterogeneous robots — machines that don’t share the same hardware, sensors or capabilities — so they can plan, divide labor and respond to changing conditions without a human micromanaging every step.
The talk framed the three problems that make this so difficult. First, autonomy: each robot needs to make sensible decisions on its own. Second, coordination: those individual decisions have to add up to coherent team behavior rather than chaos. Third, adaptability: the system can’t fall apart the moment the environment shifts or a task changes mid-mission.
APL’s answer is a scalable architecture built to support agentic behaviors across multi-robot environments. Rather than hard-coding every contingency, the approach leans on LLM-based AI agents to interpret goals, reason about the situation and translate intent into action distributed across the team. It’s the same conceptual leap that’s reshaping software assistants — but applied to physical machines that have to move, sense and cooperate in the real world.
Crucially, this wasn’t a slide-deck fantasy. The presentation included demonstrations of the approach running on actual hardware, with a heterogeneous team of robots working together rather than a single tidy prototype in a lab. That distinction matters: simulations rarely capture the awkward realities of latency, sensor noise and robots that interpret the same instruction slightly differently.
The session covered:
- An introduction to LLM-based AI agents and how they operate
- A method for applying those agents to collaborative robotic teams
- Live hardware demonstrations with a mixed team of robots
- Lessons learned so far, plus the road ahead
APL was refreshingly candid about the hard parts. The talk closed on the challenges encountered during development and the practical lessons pulled from ongoing research — the kind of honest post-mortem that’s often missing from AI hype cycles.
None of this is a product you can buy. Agentic AI for robot teams is still very much an R&D frontier, and APL’s work sits at the early, experimental end of it. But it’s a useful signal of where multi-robot autonomy is heading: away from rigid, pre-programmed choreography and toward teams that can reason about their tasks and figure out the details themselves. The accompanying whitepaper is available as a free download for anyone who wants the deeper technical picture.
