MimicX: Policy-in-the-Loop Supervision Refinement for Video-Driven Humanoid Motion Tracking
Abstract
MimicX improves humanoid motion tracking by using execution feedback to refine video-driven policies, reducing tracking errors and extending robust execution horizons.
Human videos provide rich motion targets for humanoid learning, yet visually plausible references can still produce persistent failures under physics-based execution. These failures reveal where training supervision should change. We present MimicX, a policy-in-the-loop framework that uses execution feedback to refine video-driven humanoid motion tracking. Starting from reconstructed and retargeted motion, MimicX localizes difficult transitions and affected body regions, then jointly adapts tracking objectives and the reset curriculum for policy continuation. Repeated rollout verification selects execution-priority improvements subject to tracking guards. Across four core video tasks, MimicX consistently improves tracking accuracy and Robust Execution Horizon relative to the Fixed Reference baseline. Task-averaged results show a 25.7% reduction in body-tracking error and a 255.6% increase in execution horizon. Additional video, motion-reference, and collision-scene studies evaluate the method beyond the core tasks, while MimicX-HLoop accelerates feedback through heterogeneous execution. Overall, MimicX turns policy failure into actionable supervision for deciding what to refine and which refinement to retain.
Community
🚀 MimicX: Turning Execution Failures into Better Humanoid Learning
A human video shows what a motion should look like. A physics-based rollout reveals what a robot still needs to learn. MimicX connects the two.
Rather than repeatedly training against unchanged supervision, MimicX uses the robot's own execution feedback to decide where to focus, what to refine, and which updates to keep.
🔁 From failure to refinement
- Diagnose: locate difficult transitions and body-specific tracking errors in policy rollouts.
- Refine: adapt task-aware tracking objectives and training curricula, then continue policy learning.
- Verify: retain improvements through repeated execution checks, not higher training rewards alone.
📊 Better tracking. Longer execution.
Across four video-driven humanoid tasks, our controlled experiments show 25.7% lower body-tracking error and 255.6% longer Robust Execution Horizons, averaged across tasks relative to Fixed Reference training. Further studies cover additional motions and collision-rich environments. MimicX-HLoop overlaps GPU rollout and CPU diagnosis to accelerate feedback while preserving refinement decisions.
🎬 Four skills, five synchronized views
The video brings together Tennis Swing, Football Juggling, Dance Sequence, and Kung Fu, each showing the human input alongside Fixed Reference, BeyondMimic, SONIC, and MimicX. These are the latest selected website recordings; the numerical results above come from the paper's controlled study.
Explore MimicX
🎬 Project page & videos · 📄 Paper
💻 Code · 🤗 Pretrained policies · 📦 Motion assets & data
Code, checkpoints, motion assets, and reproduction instructions are public. We'd love to hear which skills you would like to see this loop tackle next.
More videos: from human motion to robot execution
Explore more of MimicX below: full input-to-policy sequences, collision-scene execution, and detailed motion visualizations. The four-task, five-method comparison reel is in the main gallery above.
Four skills, from input to execution
Tennis Swing, Football Juggling, Dance Sequence, and Kung Fu: human demonstrations alongside recorded MimicX execution.
Execution in collision scenes
Human source clips and learned Forest Traversal and Stair Ascent execution in task-equivalent scenes.
Football motion in a registered stadium
Recorded robot motion placed in the stadium scene, with presentation props.
Foot-ball contact, in slow motion
A physical-contact excerpt driven by recorded robot motion, shown at one-quarter speed.
Parkour execution
The selected locally retimed Parkour recording.
The video-to-policy loop
Repeated verification and HLoop
Kung Fu: video-to-execution stages
Tennis motion panorama
Football motion panorama
Tennis in the court scene
Football in the stadium scene
Five collision-scene examples
Source and recorded execution snapshots across five scenes.
Project page and videos | Code | Pretrained policies | Motion assets
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- PASSAGE: Scaling Scene-Aligned Motion Learning for Perceptive Humanoid Traversal in Cluttered Environments (2026)
- KungfuAthleteBot: learning high-dynamic humanoid motion from video with unified robust recovery (2026)
- FlowHMR: Physically Plausible Motion Capture from Video (2026)
- SimVTA: Simulation-Grounded Video-Trajectory Alignment for Driving World Action Models (2026)
- PredActor: Predictive Action Diffusion for Steerable Onboard Humanoid Control (2026)
- iGPC: Generative Motion Priors for Object-Aware Humanoid Interaction (2026)
- Direct Experience World-Model Optimization: Learning the World Beyond Action Imitation (2026)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend
Get this paper in your agent:
hf papers read 2610.09055 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 1
Datasets citing this paper 1
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper