Morphmimic

MorphMimic is a morphology-conditioned world model and policy sharing system for humanoid robots running in simulation.

MorphMimic trains motion-tracking policies for humanoid robots by backpropagating through a learned, differentiable world model. It supports the Unitree G1, AgiBot X2 Ultra, and Booster T2 robots simulated in mjlab using MuJoCo Warp. Users can share world models, policies, or both between different robots by mapping them to a 29-joint layout using morphology vectors. The system features motion and five-point goal interfaces, supporting position-free tracking via IMU and joint encoders without requiring explicit state estimation. Deployment runs outside the training stack via a policy process using NumPy and ONNXRuntime communicating with a robot process over ZMQ.

Morphmimic at a glance

Best for: Researcher
Runs on: Linux, API, self-hosted, open source
Pricing: Free
Released: October 2026
Category: AI agent tool
Website: github.com/lzyang2000/morphmimic

Key features

Trains motion-tracking policies using differentiable world models
Shares policies and world models across multiple humanoid robots
Supports Unitree G1, AgiBot X2 Ultra, and Booster T2 robots
Provides motion and five-point goal interfaces
Supports position-free tracking using IMU and joint encoders
Includes anchor mode to correct drift from reference paths
Exports trained runs to ONNX for deployment
Supports teleoperation using PICO 4 headsets and motion trackers

Use it for

Training motion-tracking policies for humanoid robots
Sharing learned behaviors across different robot morphologies
Running sim2sim and sim2real robot deployment
Teleoperating humanoids with VR headsets and ankle trackers

Alternatives to Morphmimic

Supermimic Vibed — Trains a motion tracker for 40 LAFAN1 clips on a single GPU in 1.5 hours.
LoongForge-VLDA — Achieves high success rates on simulation benchmarks and real robots.
AgentGarten — Runs executable code worlds with real-time neural rendering for agents.
EvoAgentX — It automatically builds, evaluates, and evolves multi-agent workflows.
Resimind — It uses independent verification and residual-guided self-correction to grow knowledge outside model weights.
RoboHarness — It outperforms vision-language-action models without requiring robot training data.
GOD — Real-time control and intervention for agent simulations.
Reflexio — It autonomously extracts lessons from interactions to permanently correct agent behavior.
SimpleMem — Provides semantic lossless compression to reduce token consumption and maintain long-term context.
Pub Street Fighter AI — It trains a convolutional neural network on all 12 arcade levels using RGB game pixels.

All Morphmimic alternatives