One physics policy for digital humans and Unitree G1 robots

What it does

InterMimic is a research framework for physics-based whole-body control during human-object interactions. It trains a single teacher policy via reinforcement learning in NVIDIA Isaac Gym to handle dynamic, heterogeneous objects, then distills it into a student policy. The same pipeline drives both SMPL-X human avatars and Unitree G1 humanoid robots in simulation.

The interesting bit

Instead of crafting separate controllers for each body type, the project learns one policy from motion-capture data and applies it directly to both digital humans and humanoid robots. It also bridges NVIDIA’s older Isaac Gym and newer Isaac Lab, letting checkpoints trained in one environment run in the other.

Key highlights

• One unified policy covers diverse full-body interactions with varied objects.
• Supports both SMPL-X characters and Unitree G1 with three-finger dexterous hands.
• Teacher-student distillation with optional transformer-based student architecture.
• Multi-GPU distributed training support.
• Isaac Gym checkpoints compatible with IsaacLab inference.

Caveats

• The provided dataset build includes minor fixes, so reproduced results may deviate slightly from the paper.
• The codebase is tightly coupled to the NVIDIA Isaac ecosystem and requires separate environment management for Isaac Gym and Isaac Lab.

Verdict

Worth exploring if you research humanoid locomotion, manipulation, or physics-based character animation. Skip it if you need a hardware-ready robot stack or a dataset-agnostic training library.