Agentic AI for robotics: why multi-agent systems are the key to the next ChatGPT moment for robots

Agentic AI for robotics: why multi-agent systems are the key to the next ChatGPT moment for robots

🔎 The ChatGPT moment for robots won't happen as we hoped

May 2026 changes the game. IEEE Spectrum publishes a major analysis on agentic AI for robot teams, showing that the architecture that will dominate general robotics is not a single, monolithic model. It is a system of coordination, reasoning, and autonomous planning. The parallel with 2022 is striking: before ChatGPT, LLMs existed but lacked an architecture capable of making them truly useful. Today, robots exist, but they lack this "orchestrator brain" that agentic AI is beginning to provide.

Meanwhile, concrete signals are piling up. Physical Intelligence unveils π0.7, a model that recomposes learned skills to solve never-before-seen tasks. Japan Airlines deploys Unitree G1 humanoid robots at Haneda Airport. Jensen Huang decrees that Physical AI has arrived. The common denominator? None of these projects rely on a robot programmed line by line. All use agents capable of reasoning, planning, using tools, and learning from their results.

The stakes go beyond robotics. It is confirmation that agentic AI is no longer a lab concept but an industrial architecture redefining what physical machines can accomplish.

The Essentials

• IEEE Spectrum (May 2026) identifies agentic AI as the winning architecture for general robotics, replacing pre-programmed scripts with autonomous reasoning and coordination systems.
• Physical Intelligence's π0.7 demonstrates compositional generalization in robotics: a model that combines partial skills to solve novel tasks, exactly as an LLM reassembles fragments of text.
• Johns Hopkins APL publishes a functional LLM agent architecture applied to heterogeneous robotic teams on real hardware, validating feasibility outside of simulation.
• Japan Airlines is testing Unitree G1 robots at Haneda until 2028, proving that human-robot collaboration in critical environments is no longer experimental.
• NVIDIA is pushing the Cosmos model to generate synthetic training data, and Jensen Huang asserts that "every industrial company will become a robotics company."

Key tools and models of the ecosystem

π0.7 : the model that proves robots can generalize

π0.7 doesn't just execute tasks. It recomposes them.

Physical Intelligence presents π0.7 (arXiv, April 2026) as a VLA (Vision-Language-Action) model that shows "early signs of compositional generalization." Concretely? The model used an air fryer it had barely been trained on with 95% success. It had never learned this specific task. It combined partial skills — grasping, turning a knob, following a natural language instruction — to produce a new behavior.

It's the same mechanism as an LLM that has never seen a specific sentence but generates it correctly by recombining learned patterns. Except here, the output isn't text: it's a physical action in the real world. The Decoder notes that the researchers explicitly describe this approach as analogous to how a language model reassembles fragments of text.

But Physical Intelligence remains honest about the limitations. As reported by Recul.ai, the team admits that generalization depends on the human ability to articulate the task well: "It's on us. Not being good at prompt engineering." The quality of the natural language instruction directly determines the quality of the physical action. It's not a bug, it's a structural feature of the VLA approach.

MicroMatrix nicely summarizes the shift: robotics is moving from memorizing tasks to remixing partial knowledge. This is exactly the transition that occurred in NLP between 2020 and 2022.

Multi-agent architecture: why a single model is not enough

A generalist robot is not a robot with a big model. It's a robot with a system.

This is the core message of the IEEE Spectrum (May 2026): agentic AI for multi-robot systems does not replace control models. It sits on top as a layer of reasoning, planning, and coordination. An LLM agent like GPT-5.5 does not directly control the motors. It breaks down a complex task into subtasks, assigns the subtasks to specialized robots or modules, monitors the results, and adapts the plan in real time.

The approach radically changes the way a robotic fleet is designed. Instead of programming each robot with fixed instructions for every possible scenario, the team is equipped with a "coordination brain" that reasons about the situation. RobotDevDiary points out that this research area is identified by the IEEE as the most promising for heterogeneous multi-robot systems.

The connection with multi-agent systems in software AI is direct. The same principles of task decomposition, inter-agent communication, and feedback loop apply. The difference: the agents execute in the physical world, not in a terminal. This physical constraint makes error tolerance and contingent planning much more critical.

The winning architecture looks like this: a high-level agent (like Claude Opus 4.7 or GPT-5.5) for strategic reasoning, VLA models like π0.7 for low-level motor control, and an inter-agent coordination layer for teams. It is multi-stream processing applied to the physical world.

Johns Hopkins APL : validation on real hardware

Simulated demonstrations are worthless without physical proof. Johns Hopkins APL provides this proof.

The architecture presented by Johns Hopkins APL (via Xeber) applies LLM-based AI agents to heterogeneous robotic teams with demonstrations on real hardware, not in simulation. This is a crucial detail. The majority of papers on agentic robotics remain in Gazebo or Isaac Sim. Johns Hopkins steps out of the lab.

The architecture relies on hierarchical decomposition: a "supervisor" agent receives the objective in natural language, breaks it down into an action plan, and distributes subtasks to "executor" agents that each control a physical robot. Each executor agent reports back observations and results. The supervisor readjusts the plan if a sub-objective fails or if the environment changes.

What makes this approach powerful is that it works with heterogeneous robots. There is no need for a uniform fleet. A KUKA arm, a drone, a wheeled mobile robot — all can participate in the same mission because coordination occurs at the semantic level (natural language), not at the protocol level. This is a paradigm shift for industrial robotic integration.

For companies that want to explore these architectures without investing in expensive hardware, the path of open source AI with Ollama locally offers a realistic testing ground for the coordination layer.

Haneda Airport : when humanoid robots enter production

Theory meets reality in Tokyo.

Japan Airlines is testing Unitree G1 humanoid robots at Haneda Airport for baggage handling. The trial runs until 2028 and explicitly focuses on safe human-robot collaboration in a high-density pedestrian environment. NewsGab reports that the primary goal is to solve the labor shortages hitting the Japanese aviation sector.

This deployment is significant for several reasons. First, it is not driven by a research lab but by an airline with real operational KPIs. Second, it puts humanoids in direct contact with the public, not in a closed workshop. Finally, the duration of the test (until 2028) indicates that JAL is not doing PR but a serious evaluation of profitability and reliability.

The use case is "simple" — carrying luggage — but the environment is chaotic. A running child, a poorly parked cart, a falling suitcase: so many unpredictable situations that agentic architectures are designed to handle, where a pre-programmed script would come to a dead stop. This type of deployment reminds us that precision manufacturing is no longer the only field where robots prove their value.

NVIDIA and the "Big Bang" of Physical AI

Jensen Huang doesn't do things by halves.

During his presentation on agentic AI at NVIDIA, Huang declares that "Physical AI has arrived" and that "every industrial company will become a robotics company". The statement is ambitious, but it rests on a concrete element: the Cosmos model.

Cosmos does not control robots. It generates future states of the world in the form of synthetic videos. The idea is powerful: rather than having to record millions of hours of robotic data in the real world (slow, expensive, dangerous), Cosmos simulates physically plausible scenarios that robotic models can use as training data. It is synthetic data, but with a physical consistency that makes it usable for learning.

This approach solves a major bottleneck. π0.7 and similar VLA models need diverse action data to generalize. NVIDIA provides the data "well". Startups like Physical Intelligence provide the model. Agentic LLMs provide the coordination. The ecosystem completes itself.

The parallel with the history of LLMs is enlightening. The generalization of LLMs exploded when we had both the model (Transformer), the data (internet-scale) and the infrastructure (GPU). For robotics, the three pieces are now on the board.

Governance: the invisible brake that could slow everything down

The more autonomous robots become, the more urgent the question of a regulatory framework becomes.

Agentic AI governance initiated by Google and SAP precisely aims to regulate AI agents in the enterprise. But in robotics, the stakes are of a different order. A software agent that makes a mistake costs money or time. A robotic agent that makes a mistake can injure someone.

The governance of robotic multi-agent systems raises unprecedented questions. Who is responsible when a supervisor agent makes a decision that leads an executor robot to cause damage? The robot manufacturer? The LLM provider? The company that deployed the system? The human operator who wrote the initial prompt?

These questions are not theoretical. JAL's tests at Haneda take place in a public space. The architectures from Johns Hopkins APL are designed for missions involving powerful physical robots. The compositional generalization of π0.7 means by definition that the robot will do things that have not been explicitly tested.

Risk is not an argument against deployment. It is an argument for architectures where governance is integrated by design — human supervision loops, physical safeguards, explicit competence limits that the agent cannot override.

Choosing LLMs for robotic orchestration

Not all LLMs are created equal when it comes to controlling physical agents.

The agentic benchmark (June 2025) provides clear indications. OpenAI's GPT-5.5 dominates with a score of 98.2, making it the natural candidate for the high-level reasoning layer. Its ability to break down complex tasks into atomic steps is critical when each step corresponds to an irreversible physical action.

Google's Gemini 3 Pro Deep Think, with its score of 95.4, excels in multi-sensor scene analysis — useful when the robot must fuse visual, spatial, and textual data to make a decision. Anthropic's Claude Opus 4.7, at 94.3, stands out for its adaptive reasoning, which allows a plan to be adjusted mid-execution without recalculating everything.

For local or air-gapped deployments — a common case in industrial robotics for latency and security reasons — Moonshot AI's Kimi K2.6 (score 88.1, self-host) and Z.AI's GLM-5 (score 82, self-host) offer credible alternatives. The performance/autonomy trade-off is real but often acceptable for well-defined coordination tasks.

The choice of LLM depends on the level in the hierarchy. A model like GPT-5.5 for the strategic supervisor. A lightweight and fast model for the executing agents that must react in milliseconds. This strategic selection of LLMs for agents is a key skill that robotics teams must now master.

❌ Common mistakes

Mistake 1: Confusing VLA control and agentic reasoning

Many commentators treat π0.7 as a "robotic agent". This is inaccurate. π0.7 is a VLA model — it maps a visual observation and a language instruction to a motor action. It does not plan, break down tasks into subtasks, or coordinate with other robots. Agentic reasoning is a higher layer that uses VLA models as execution tools. Mixing the two levels leads to overhyped architectures that do not scale.

Mistake 2: Believing that simulation replaces the real world

NVIDIA Cosmos is a powerful tool for generating synthetic training data. But as demonstrated by the approach of Johns Hopkins APL, final validation must be done on real hardware. The "sim-to-real gap" remains a major obstacle: a robot that performs at 99% in simulation can fail at 30% in the real world because of friction, communication delays, and mechanical inaccuracies that simulation does not perfectly capture.

Mistake 3: Ignoring physical prompt engineering

Physical Intelligence made it clear: generalization depends on the quality of the human instruction. Deploying an agentic robotic system without training operators to write precise natural language instructions is like putting a V8 engine in a car without a steering wheel. "Prompt engineering" is not just an AI blogger trick — it is a critical operational skill in agentic robotics.

Mistake 4: Deploying a single model to do everything

The illusion of the single model persists. In practice, working agentic robotic systems use a specialized stack: LLM for reasoning, VLA for control, perception models for vision, prediction models for navigation. Forcing a single model to do everything guarantees mediocrity on every dimension.

❓ Frequently Asked Questions

What is compositional generalization in robotics?

It is a model's ability to combine separately learned skills to solve a never-before-seen task. π0.7 demonstrated this by manipulating an air fryer without having been trained on it, by recombining partial know-how (grasping, turning, following an instruction). It is the robotic equivalent of what LLMs do with text.

Is a single agentic robot better than a team of specialized robots?

No, it's the opposite. The winning architecture identified by IEEE Spectrum is a heterogeneous team coordinated by a high-level agent. Each robot excels in its own domain, and the agentic agent ensures they collaborate effectively. The "one robot to rule them all" approach is a myth inherited from sci-fi movies.

Can agentic LLMs run locally on a robot?

Yes, partially. Models like Kimi K2.6 (88.1) and GLM-5 (82) are designed for self-hosting. But for complex reasoning tasks requiring GPT-5.5 or Claude Opus 4.7, a cloud connection remains necessary. The trade-off between latency, bandwidth, and reasoning power is an active area of research.

Is JAL's deployment at Haneda really agentic AI?

The JAL test focuses primarily on human-robot collaboration in a public environment. The exact level of agenticity is not public. But the context — unpredictable environment, interaction with humans, real-time adaptation — is precisely where agentic architectures show their superiority over scripted approaches.

When can we expect a robotic "ChatGPT moment"?

The pieces are in place: generalizing VLA models (π0.7), powerful agentic LLMs (GPT-5.5, Claude Opus 4.7), synthetic data (Cosmos), and real-world deployments (Haneda). But the ChatGPT moment doesn't arrive when the technology is ready — it arrives when an interface makes it accessible. That layer of abstraction is still missing for robotics.

✅ Conclusion

Agentic AI is not going to improve robots. It is going to reconceptualize them — shifting from machines that execute to systems that reason, plan, and adapt. To keep up with the evolution of this convergence between software agents and physical agents, check out our monthly AI trends watch.