MorFiC: Fixing Value Miscalibration for Zero-Shot Quadruped Transfer

The result in one glance

One policy. Different bodies.

MorFiC takes a different route to generalization: improve the value function so a policy trained on one source robot can transfer efficiently, rather than relying on target-specific retraining or very large multi-embodiment training budgets.

1source robot

7simulated morphologies

2real robot platforms

≈2hsingle-source training

Efficient generalization

Not brute-force scaling. MorFiC targets a specific actor–critic failure mode — value miscalibration across bodies — enabling competitive zero-shot transfer with a compact single-source training setup.

Zero-shot cross-morphology transfer A single learned policy evaluated across seven robot morphologies in simulation.

The diagnosis

Why does zero-shot transfer break?

Across robot morphologies, the same observed state can lead to very different future returns because mass, kinematics, joint limits, and actuation change the underlying dynamics.

A shared critic is therefore pushed toward averaging incompatible value targets. That value bias flows directly into PPO advantage estimates — and can change the direction of policy updates.

V

The core idea: cross-morphology generalization is not only a policy representation problem. The value function must also stay calibrated as the body changes.

Critic calibration under morphology shift for PPO variants — Explained variance across target morphologies. Higher is better.

Method

Condition the critic multiplicatively.

MorFiC encodes a robot morphology descriptor into a latent representation shared by the actor and critic. The critic uses morphology-aware multiplicative modulation, allowing the value landscape itself to change with the embodiment instead of treating morphology as a weak additive feature.

MorFiC architecture showing morphology latent conditioning the actor and critic — MorFiC architecture. A morphology latent modulates the critic through multiplicative conditioning while also conditioning the actor.

01

Sample morphology

Physical and control parameters are randomized around the source robot during training.

02

Encode the body

A lightweight encoder maps the morphology descriptor to a shared latent representation.

03

Modulate value

Multiplicative critic conditioning captures morphology-dependent value landscapes and improves advantage calibration.

Cross-morphology transfer

Seven bodies, zero target-robot training.

Evaluation spans Go2, Go1, A1, AlienGo, Cheetah, ANYmal, and B1. Performance remains strongest near the source morphology and degrades gracefully as the embodiment moves farther out of distribution.

Seven simulated quadrupeds Same learned policy, unseen target morphologies.

Mean forward velocity versus normalized morphology distance from Go2 — Morphology-distance analysis: MorFiC preserves stronger locomotion as robot morphology shifts away from the training source.

Real hardware

Same policy weights. Go1 and Go2.

MorFiC transfers from simulation to two Unitree platforms without target-specific fine-tuning, reaching approximately 1.4–1.7 m/s on Go1 and 1.1–1.3 m/s on Go2.

Zero-shot hardware deployment Indoor and outdoor runs on Go1 and Go2.

Perturbation recovery Push disturbances on real hardware.

Efficiency + evidence

Generalization without brute-force compute.

MorFiC trains on a single source embodiment in roughly two hours while retaining strong zero-shot transfer. The critic diagnostics help explain why: better value calibration yields more reliable advantages and policy updates across morphology shift.

External baselines and critic calibration under morphology shift — Critic calibration across external baselines and target morphologies.

Training efficiency versus external baselines — MorFiC trains on one source embodiment in roughly two hours, versus substantially larger training budgets reported by scaling baselines.

1.98 m/sAlienGo forward velocityvs. additive PPO variants below 0.65 m/s in the reported comparison
~5×B1 speed gainrelative improvement reported against morphology-conditioned PPO baseline
0target fine-tuningfor the reported Go1 and Go2 hardware deployments

Full overview

MorFiC in 2 minutes 48 seconds.

The complete supplementary overview: motivation, method, diagnostics, simulation transfer, hardware deployment, and limitations.

MorFiC supplementary video

Stress tests + limitations

Where transfer still gets hard.

Performance falls as morphology moves far outside the training distribution. Real-world deployment is also sensitive to latency, state-estimation drift, and contact mismatch. The accompanying reel intentionally starts with failure cases, then shows additional successful hardware runs to give a fuller picture of deployment behavior.

Read the appendix →

Stress tests, failures & additional hardware runs Failure cases first, followed by additional successful runs.

News & coverage

One Robot, Many Bodies.

The University of Maryland Institute for Advanced Computer Studies profiled MorFiC, its single-source training setup, critic-miscalibration diagnosis, and zero-shot Go1/Go2 deployment.

UMIACS · September 18, 2026

Read the UMD feature One Robot, Many Bodies: New System Helps Four-Legged Robots Adapt ↗

Citation

Use MorFiC in your work.

If you find the method or diagnostics useful, please cite the paper.

@article{mishra2026morfic,
  title   = {MorFiC: Fixing Value Miscalibration for Zero-Shot Quadruped Transfer},
  author  = {Mishra, Prakhar and Raj, Amir Hossain and Xiao, Xuesu and Manocha, Dinesh},
  journal = {arXiv preprint arXiv:2603.14554},
  year    = {2026}
}