Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

RedirectMimic: Local Motion Content Redaction for Humanoid Whole-Body Control

Zepeng Wang, Xiaochuan Shi

November 2026

Side-by-side simulation rollouts: RedirectMimic against ZeroTrack, AntiTrack and the adapted ForgetMimic.

TL;DR Fine-tune a humanoid motion tracker so it swaps out selected arm movements while still being fed the original reference, by computing tracking rewards against a rule-edited reference. In Unitree G1 simulation, overhead-arm content drops to 2.7% of rollouts with 94.2% joint success.

What they do

A humanoid motion tracker normally reproduces whatever reference it is given. RedirectMimic edits the policy so that a specified arm movement (an overhead reach, a fast outward extension) is replaced by a chosen alternative, while the rest of the motion and the robot’s balance are kept. The deployed policy still receives the original, unedited reference; the replacement lives in its weights.

How it works

  • Paired references. An offline rule finds the unwanted movement in each clip and blends the selected arm joints toward a neutral pose (upper arm near the torso, elbow bent), with smooth transitions. The root trajectory and all other joints are untouched; body poses and velocities are recomputed by forward kinematics.
  • Input/target split. During PPO fine-tuning of a pretrained tracker, the actor observes the original reference, but BeyondMimic-style tracking rewards are computed against the edited one.
  • Edit-window sampling. Episode starts mix the base failure-weighted sampler with a pool of frames around the edits, since edited frames are sparse.
  • Baselines. ZeroTrack removes the edited arm’s tracking reward, AntiTrack rewards deviating from it, and an adapted ForgetMimic cancels tracking reward on whole clips.

Results

  • Setup: 29-joint Unitree G1 in mjlab, tracker pretrained on 7,509 retargeted AMASS clips, 149 overhead (S2) and 173 speed-defined (S1) windows, 16 rollouts each.
  • S2: content occurs in 2.7% of rollouts against 58.9% (ZeroTrack) and 32.8% (AntiTrack); joint success is 94.2% against 40.8% and 65.0%; arm error to the target is 0.079 m against 0.224 m and 0.171 m.
  • S1: all three local methods suppress the content similarly (2.0-2.3%). ZeroTrack has the highest joint success (0.897 against 0.874), but RedirectMimic has the lowest replacement error (0.143 m).
  • On 200 unedited clips per task, fall rate and tracking error stay close to the pretrained tracker (S2: 0.098 and 0.070 m against 0.095 and 0.070 m). The adapted ForgetMimic falls in every edited rollout.

Limitations

Simulation only, on known clips; physical humanoids are left to future work. The editor uses the whole clip to place transitions while the actor sees only the current frame, so transition timing under streaming references is untested. On S2 the fall rate (2.2%) is higher than AntiTrack’s (0.9%). The ForgetMimic baseline is the authors’ reimplementation under a shared protocol.

Videos

All six controllers on one overhead-arm window, with wrist height plotted over time.
Another six-controller comparison on a single edit window.

Figures

From Figure 2: RedirectMimic overview. (a) Construct the local replacement. (b) Fine-tune with original-reference obs...
From Figure 2: RedirectMimic overview. (a) Construct the local replacement. (b) Fine-tune with original-reference observations and edited-target rewards. (c) Retain the deployment interface.
From Figure 3: Stretching (a) and cup drinking (b) at core entry, mid-core, and after the edit. The blue arm shows th...
From Figure 3: Stretching (a) and cup drinking (b) at core entry, mid-core, and after the edit. The blue arm shows the original reference and the orange arm shows the edited target.
From Figure 4: Content suppression and replacement fidelity. The top and bottom rows show S2 (overhead arm) and S1 (f...
From Figure 4: Content suppression and replacement fidelity. The top and bottom rows show S2 (overhead arm) and S1 (fast outward arm). The left panels show per-frame content probability around the edit core. The right panels show mean elbow-wrist distance to the edited target during entry, core, and exit.
Abstract

Humanoid motion interfaces should accommodate application-specific movement constraints while retaining useful reference behavior. Local content redaction requires a controller to replace selected movements even when it continues to receive the original reference. Removing the tracking incentive for those movements leaves their replacement unspecified. We introduce RedirectMimic, 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 task, 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 the 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/.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.