Teaching a humanoid robot to pick up a cup, walk across a cluttered room, and open a door—tasks trivial for humans—has been one of robotics’ hardest problems. Motion capture data exists in abundance. The challenge: making that data work on robots with different bodies, in real environments, at scale.
InterMimicGen, released October 5 on Hugging Face, attacks this problem with a self-improving flywheel that turns a single human demonstration into an expanding dataset of robot skills.
The Core Idea
The system takes motion-capture recordings of humans interacting with objects—reaching, grasping, walking while carrying—and “retargets” them onto humanoid robots with dexterous hands. This isn’t simple replay. Humans and robots have different proportions, joint configurations, and centers of gravity.
InterMimicGen trains a generalist tracker in simulation that watches a motion and predicts what the robot’s joints should do. Then it runs a self-evolving loop: generate small task-preserving edits to the motion, test them with the tracker, and keep only the variations where the tracker still succeeds.
Each successful variant becomes new training data. The tracker gets better. It generates more diverse motions. The cycle compounds.
What It Achieves
The paper reports results across nine loco-manipulation tasks in simulation. The self-evolving flywheel produces a 340% increase in successful motion variants compared to a single-pass retargeting baseline. The generalist tracker achieves 91.2% success rate on novel motions it hasn’t seen during training.
Crucially, the system works across different humanoid platforms. The same motion library can be retargeted to robots with different arm lengths, torso heights, and hand designs—without manual per-robot tuning.
Why It Matters
General-purpose humanoid robots need massive, diverse skill libraries. Manual motion capture is expensive. Teleoperation doesn’t scale. InterMimicGen shows how a single demonstration can bootstrap an entire curriculum of robot behaviors, with the system discovering variations a human teacher wouldn’t think to try.
The code and trained models are available on Hugging Face, with the full evaluation benchmark released for the community to build on.