RedirectMimic: Local Motion Content Redaction for Humanoid Whole-Body Control
RedirectMimic: Local Motion Content Redaction for Humanoid Whole-Body Control
This paper presents RedirectMimic, a technique that enables local motion content redaction for humanoid whole-body control by separating the reference observed by the actor from the target used to train its execution. The method uses a rule-generated edited reference to guide tracking rewards during proximal policy optimization while keeping deployed observation and action interfaces unchanged. Evaluated on Unitree G1 simulation with a tracker pretrained on 7,509 motion clips, RedirectMimic reduces core-window content occurrence to 2.7% on the overhead-arm task (compared to 58.9% for reward zeroing and 32.8% for anti-tracking), achieves 94.2% core-window joint success, combines content suppression with fall-free execution and local retention, and maintains low replacement error while tracking unedited motions with accuracy comparable to the pretrained controller.
The authors introduce RedirectMimic, a method for local motion content redaction in humanoid whole-body control that separates the reference observed by the actor from the target used to train execution, evaluated on Unitree G1 simulation with reported core-window content occurrence of 2.7% and joint success of 94.2%.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S3
Reported numbers
motion clips count
7,509
View original evidence
rectMimic, which separates the reference observed by the actor from the target used to train its execution. A rule-generated edited reference guides tracking rewards during proximal policy optimization, while the deployed observation and action interfaces remain unchanged. We evaluate pose- and speed-defined arm editing in Unitree G1 simulation using a tracker pretrained on 7,509 motion clips. On the overhead-arm tas
Open source S3
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
rectMimic, which separates the reference observed by the actor from the target used to train its execution. A rule-generated edited reference guides tracking rewards during proximal policy optimization, while the deployed observation and action interfaces remain unchanged. We evaluate pose- and speed-defined arm editing in Unitree G1 simulation using a tracker pretrained on 7,509 motion clips. On the overhead-arm tas
Open source S3
k, RedirectMimic reduces core-window content occurrence to 2.7%, compared with 58.9% for reward zeroing and 32.8% for anti-tracking. It achieves 94.2% core-window joint success, combining content suppression, fall-free execution, and local retention. On the speed-defined task, fast outward arm movements occurred at similarly low rates under all three local editing methods. Across both tasks, RedirectMimic achieved th
Open source S4
e lowest replacement error among these methods and tracked unedited motions with accuracy comparable to that of the pretrained controller. Project page: https://xiaopeng-whu.github.io/RedirectMimic/.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org