Moving a large object with several robots sounds straightforward until the robots themselves cannot get into the positions needed to push it. Researchers at Carnegie Mellon University have developed a planning framework for robot teams designed for exactly that problem. LAMP, short for Long-Horizon Adaptive Manipulation Planning, combines a learned generative model with classical search so teams of robots can plan object movement and their own movement together.

The system targets cluttered environments where a short path for the object may be useless because shelves, walls or other obstacles block the robots from reaching the required contact points. The research is being presented at IROS 2026.

Robot teams and the object have to be planned as one problem

A conventional planner might first decide where an object should travel and only then work out how the robots can push it along that route. In an open room that can be enough. In a crowded warehouse, a seemingly perfect object path may leave no space for the robots to stand, turn or change sides.

LAMP tackles the coupled problem. It reasons about contact formations, robot motion, collision avoidance and the object’s path over a longer sequence of actions. The researchers describe that combined search space as difficult because it expands rapidly as more robots, more obstacles and more steps are added.

The framework uses a learned generative manipulation model to propose useful local movements, then applies systematic planning to connect them over a longer horizon. One version, LAMP-A*, searches the coupled robot-object space using an A*-style planner. A second version, LAMP-Lazy, delays expensive checks until they are actually needed and can replan as new information arrives.

That mixture of learning and classical planning is notable. The learned model helps generate plausible actions without forcing the entire problem into an end-to-end neural policy. The planner preserves an explicit search structure for the longer task.

The warehouse-style demonstration gets harder as it succeeds

Carnegie Mellon’s description of the project highlights a demonstration in which multiple robots transported 12 objects one at a time to spell the letters IROS.

The visual trick is useful because the workspace becomes more difficult as the task progresses. Each placed object changes the free space available for later movements. The robots therefore cannot treat the environment as a static puzzle solved once at the beginning.

CMU says LAMP-Lazy can update the plan using new information from the robots and their surroundings. That kind of feedback is essential outside simulation because physical manipulation never follows a mathematical plan perfectly. Wheels slip, objects rotate slightly differently and contact forces change.

The research paper reports experiments in challenging simulated environments where the approach solved long-horizon tasks in densely cluttered scenes that the comparison methods could not handle. That is promising evidence for the planning method, but the paper should not be read as proof that warehouses can simply deploy LAMP tomorrow. Simulation and controlled robot demonstrations are steps towards operational reliability, not the same thing as it.

Multi-robot manipulation has a coordination tax

Adding robots can provide more force, stability or control, but every extra robot also introduces another moving body that must avoid collisions and reach the correct place at the correct time.

This is the coordination tax in multi-robot work. Two robots can accomplish something one robot cannot, while also creating a planning problem that is much larger than twice as hard. Their motions are coupled through the object they are manipulating, so one robot’s useful position can block another’s route.

LAMP’s value is its attempt to treat coordination as part of the planning problem from the start. The robot team is not a collection of independent movers that happen to touch the same object. It is a joint system whose feasible options depend on everybody’s position.

LiveAIWire has looked at ways of transferring robot skills between different bodies. LAMP addresses a complementary challenge: even capable individual robots need a shared plan when the task itself requires cooperation.

Learning is being used to shrink search rather than replace it

Robotics increasingly mixes learned models with older planning and control techniques. That can sound less fashionable than an end-to-end model that learns everything, but physical systems benefit from explicit structure.

A learned local model can capture complicated pushing behaviour or suggest promising contacts. A search algorithm can then reason about sequence, feasibility and changing geometry. The combination can be easier to inspect than a single policy that maps perception directly to action.

This hybrid approach is also relevant to safety. A robot planning around people, shelves and expensive equipment needs more than a plausible action. It needs to maintain constraints over a chain of actions and recover when the world does not match the prediction.

LiveAIWire’s coverage of robot safety refusals shows the other side of physical AI: knowing when an action should not be attempted is part of useful autonomy. Better planning can reduce the number of situations where a robot has to choose between making progress and entering a risky configuration.

Warehouses are an obvious application, but not the only one

CMU points to warehouses because they contain exactly the kind of clutter and repeated object movement that makes multi-robot manipulation attractive. A team could move items too awkward for one robot, navigate around fixed infrastructure and adapt as inventory changes.

The same planning problem appears elsewhere. Construction sites, manufacturing cells, disaster-response environments and laboratories can all require several machines to manipulate something while space is limited.

The common feature is not the industry. It is the need to coordinate contact, motion and access over a long sequence rather than solving one push at a time.

LiveAIWire has also covered the energy limits of mobile robots. In real deployments, planning quality affects energy as well as success. Unnecessary repositioning, failed attempts and long detours all consume power and time.

The interesting advance is making cooperation part of the route

Robotics demonstrations often focus on whether a machine can move an object from A to B. LAMP makes the hidden problem visible: the route is not valid unless the robots can physically arrange themselves to execute it.

That sounds obvious after it is stated, but it changes how the planner must reason. The object does not move through an empty geometric world. It moves because several physical agents repeatedly find reachable contact positions around it.

If LAMP and related approaches scale beyond research environments, robot teams could become useful in places where one large specialised machine is currently required. The advantage would come not from making every robot stronger, but from making the group better at finding a sequence of moves that remains feasible as the space closes around them.

Long-horizon planning matters because one good move can create a bad future

A robot can make a locally sensible decision and still leave the team trapped later. Pushing an object through a narrow opening may move it closer to the destination while removing the space another robot needs to reach the next contact point. Multi-step manipulation therefore requires the planner to think about future access as well as immediate progress.

That is the reason the long-horizon part of LAMP matters. The system considers the motion of the object and the robots together over a sequence of actions. It is trying to avoid plans that look feasible in geometry but break down when the machines actually need to position themselves around the object.

LAMP-Lazy adds another practical idea: not every possible movement has to be fully checked in advance. The system can defer some verification until a movement becomes relevant and then use new information to update the plan. That can save effort in problems where many hypothetical branches will never be used.

Real warehouses would add uncertainty the demonstration can only approximate

The research tests the planning framework in controlled environments, including crowded simulated settings and a physical demonstration. A working warehouse would add moving people, changing inventory, imperfect localisation, wheel slip, blocked aisles and equipment that was not in exactly the expected place.

Those complications do not diminish the result, but they explain why a successful planning paper is not the same as a ready-made warehouse product. Robust deployment would require perception and control systems that continually update the planner with reliable information, plus safe behaviour when the planned action can no longer be executed.

The same issue appears whenever several autonomous machines share a workspace. Coordination is not only about avoiding collisions between robots. They must avoid making each other’s future tasks impossible.

Better coordination could let simpler robots tackle harder jobs

One attraction of multi-robot manipulation is that capability can come from cooperation rather than from building one machine large enough to handle every object alone. Smaller robots can potentially be reassigned across jobs and combined when a task needs more stability or control.

That flexibility only works if planning can handle the added coordination cost. As the team grows, so does the number of possible positions and interactions that have to be considered. LAMP is an attempt to keep the useful part of cooperation while planning through that complexity.

The broader lesson is that robot intelligence is increasingly about sequences rather than isolated skills. A machine may know how to push, grasp or navigate, yet still fail a real task because those abilities are used in the wrong order. Systems such as LAMP focus on the orchestration layer: deciding which feasible action should happen now so that the next necessary action remains feasible later.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.