Robot skill transfer usually runs into an awkward physical problem: software trained to control one machine may fail when the legs, weight, joints or proportions change. MorFiC, a research system accepted for CoRL 2026, tackles that problem by training a locomotion policy on one quadruped and moving it to different four-legged robot bodies without fine-tuning the policy for each new machine.

The researchers tested the approach across several quadruped platforms and demonstrated real zero-shot transfer on Unitree Go1 and Go2 robots. The claim is narrower than a universal robot brain. MorFiC is about locomotion among related four-legged morphologies, and performance still depends on how different the target body is. But the direction is important: useful robot behaviour becomes much easier to deploy if every new body does not require a fresh learning programme.

How robot skill transfer works across different bodies

Robotic locomotion policies learn a relationship between what the robot senses and how its motors should respond. That relationship is tied to the body. A longer leg changes leverage. A heavier torso changes balance. Different joint limits change which movements are physically possible. A policy that works beautifully on one robot can therefore become unstable when copied to another.

MorFiC tries to separate the useful structure of locomotion from the exact source body. During training, the system exposes the policy to variations in morphology in simulation rather than allowing it to become perfectly specialised for one fixed robot. The aim is to teach a control strategy that remains useful when the physical parameters shift.

The University of Maryland’s September account of the work says the policy can be trained on a single robot and then transferred to other quadrupeds without retraining for each target. That is the practical attraction. The expensive part of learning can potentially be reused rather than repeated.

The body still matters even when the policy travels

It would be easy to overread the result as evidence that robot intelligence has become hardware-independent. It has not. MorFiC works because the target machines share a broad quadruped structure and because the training process anticipates variation in that structure.

A four-legged controller is not suddenly ready for a humanoid, a wheeled warehouse machine or a robotic arm. Even within quadrupeds, sufficiently large differences in mass distribution, actuator strength or geometry can create a transfer gap that the learned policy cannot bridge.

That limitation is useful rather than disappointing. Robotics needs methods that generalise one step further than the current system, not magical claims that embodiment no longer matters. LiveAIWire’s coverage of the race to build capable robotic hands illustrates the same principle from another angle. Intelligence has to meet the mechanical reality of the body that acts on it.

Real Go1 and Go2 deployments make the result more interesting

Simulation is an essential part of modern robotics because it allows huge numbers of trials without breaking hardware. It also creates a familiar problem: a controller can exploit simplifications in the simulated world that disappear when the robot touches a real floor.

MorFiC was therefore not assessed only on virtual machines. The researchers deployed the transferred policy on real Unitree Go1 and Go2 quadrupeds without fine-tuning. That does not prove reliability in every environment, but it shows the core transfer idea surviving the simulation-to-reality step on physical robots.

The paper also reports stronger transfer speeds than a morphology-conditioned PPO comparison on several target platforms. Those benchmark gains matter to specialists, but the more useful general point is that the target body did not need its own full training process before it could walk under the transferred policy.

Why reusable locomotion could change robot economics

Hardware diversity is expensive for software teams. A company deploying several robot models can end up maintaining separate policies, tuning procedures and testing pipelines for machines that perform broadly similar tasks. Every hardware revision risks restarting part of that work.

A transferable locomotion layer could make the software stack more portable. That would not remove engineering work around sensors, safety limits, payloads and task-specific behaviour, but it could reduce the amount of relearning required just because a robot’s dimensions changed.

The idea also matters for research. Laboratories often work on different platforms, which makes it harder to reproduce results. A method that survives body changes gives researchers a better chance of testing the same control concept across several machines rather than treating every robot as its own isolated ecosystem.

Transfer does not solve robot safety

A robot that can reuse a walking policy is not automatically a robot that knows when it should walk, where it should go or what actions are safe around people. Those questions sit above locomotion in the control stack.

LiveAIWire recently examined how robot AI systems handle unsafe instructions and refusals. MorFiC addresses a different layer. It is concerned with making a body move effectively, not deciding whether the requested action is appropriate.

That distinction will become increasingly important as reusable robot models spread. Portability is valuable precisely because it allows capabilities to move more easily between machines. Safety testing, physical limits and environment-specific validation therefore have to travel with equal care.

The important shift is from one robot to a family of bodies

Robotics has often progressed by producing impressive skills on one carefully prepared machine. MorFiC asks a more practical question: how much of that skill survives when the machine changes?

The answer is not “all of it”. The work remains focused on quadruped locomotion, and larger morphological differences can still break transfer. But the real-world Go1 and Go2 demonstrations show that a learned walking strategy can be less attached to one exact body than conventional training would suggest.

If that principle extends to broader robot families, future control systems may be developed more like software platforms. A capability could be trained once, adapted through a known physical interface and deployed across related hardware. The robot would still need a body. It might simply stop needing a completely new education every time that body changes.

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.