A University of Maryland and George Mason University team has transferred one simulated Unitree Go2 locomotion policy across seven quadruped platforms by changing how reinforcement learning evaluates an action. The MorFiC framework conditions its critic on the robot's physical form, then runs on new bodies without retraining; the team also deployed the same policy weights on real Go1 and Go2 robots.
The method targets the training feedback used in actor-critic reinforcement learning. The actor chooses a movement, while the critic predicts the rewards the robot can earn from its current situation. Training compares observed returns with that prediction to decide which actions to encourage. A critic that treats different bodies alike can give misleading feedback: the same pose or command may work differently when masses, joint limits and motors change.
MorFiC, short for Morphology-aware FiLM Critic, turns those physical properties into a numerical description of the robot. It uses that description to scale and shift signals inside the critic, so the same network can judge a situation differently for a light robot and a heavy one. The rest of the training procedure stays the same. The change is therefore in how the robot's performance is evaluated during learning, not a separate movement policy retrained for every body.
The researchers trained on a single Go2 template with randomized physical parameters in simulation and evaluated the policy on Go2, Go1, A1, AlienGo, MIT Mini Cheetah, B1 and ANYmal C models. Internal controls used the same training setup but either omitted morphology information, sent it only to the actor, or concatenated it to both actor and critic. The paper measured forward speed and survival together because a robot can remain upright while barely moving.
In that internal comparison, MorFiC reached 1.98 meters per second on AlienGo while the additive actor-critic variants stayed below 0.65 meters per second. Against four external cross-morphology methods, it posted the highest forward speed on five of seven simulated robots. The result was not universal: another method reached 2.86 meters per second on the heavier B1, where MorFiC managed only 0.21 meters per second and did not produce an effective gait.
The hardware test was narrower. Across 10 Go1 trials indoors and outdoors, the policy ran at roughly 1.4 to 1.7 meters per second and completed more than 80% of runs without freezing or gait failure. On Go2, it reached about 1.1 to 1.3 meters per second. The team reported latency, state-estimation drift, foot slip and uneven steps, with the Go2 deployment showing slower movement and more failures.
MorFiC therefore reduces one source of transfer error without making robot shape irrelevant. Performance fell as a target's normalized morphology moved farther from the Go2 training body, especially for B1 and ANYmal. Expanding the randomization range made the learned policy more cautious rather than fixing distant transfer, and the study did not test bipeds or humanoids.