Humanoid motion-tracking policies are typically developed under nominal actuation, yet actuators can weaken or fail over a robot's service life. A local actuator impairment can affect whole-body tracking and require coordinated compensation from intact body parts. We therefore ask whether organizing the control representation around body parts provides a useful inductive bias for fault-tolerant tracking. We present FailureMimic, a framework that distills a whole-body tracking policy into a part-wise latent motion model and adapts the frozen model with a residual policy that coordinates across body parts. The residual policy is not given the identity of the impaired joints. In simulation on a Unitree G1, FailureMimic achieves higher fall-free survival than monolithic and non-part-aware baselines under passive leg impairments and systematic pushes; its margin grows with impairment severity and persists at severities not encountered during training. The same structure also serves the opposite demand: because body parts stay separable, it tracks novel combinations of upper- and lower-body motions. On the physical G1, FailureMimic autonomously performs loco-manipulation with passive or locked joints and tracks stitched motions.
Overview of the three stages. Stage 1: a whole-body expert tracks reference motions g, emitting joint-position targets. Stage 2: the expert is distilled into a part-wise latent prior p(z | s), aligned by a divergence term with a training-time posterior. Stage 3: the prior and decoder are frozen, and a body-part Transformer maps the state and the sparse five-point command to a residual. Colors mark the five trunk and limb groups.
Part-wise architecture. The command, proprioceptive state, and attitude signals are tokenized by body part into five tokens; a Transformer encoder mixes them by self-attention, and zero-initialized per-part heads emit the latent residual. The residual shifts the frozen prior mean, and each latent slice is decoded by an independent branch driving only its own group's joints. Snowflakes mark frozen modules; the bolt marks the trained residual generator.
The G1 has 29 actuated joints, grouped into the trunk, two arms and two legs, the same five-part partition the latent model is factorized over. A damped joint has its stiffness set to zero: it gives no position feedback and contributes only dissipative torque. A locked joint is held at its episode-start configuration. The policy is never told which joints are affected.
The ringed joints below are only the ones demonstrated on this page. They are not a limit of the method. Any joint can be damped or locked, and the framework trains the same way for it. Hover a ringed joint to see where it appears; click to jump to that clip.
Loading model…
drag to rotate · scroll to zoom
Hover or select a joint.
Box carrying
Box carrying with right-wrist damping
Box carrying with right hip-roll damping
Box carrying with left-elbow locking
Box pushing
Box pushing with right ankle-roll damping
Damping the left ankle-roll joint
Damping the right hip-yaw joint
Damping the right ankle-roll joint
Damping the left ankle-pitch joint
Locking the left knee
Locking the left hip-roll joint
Each policy is trained and evaluated at the same impairment severity, then pushed. The part-structured residual holds up as the compensation demanded of it grows: against the unfactorized baseline its advantage widens from 6.0 points at one impaired joint to 30.6 at three.
Fall-free survival under push
71 motions · 4096 trials per cell
Tracking novel upper/lower-body combinations
500 stitched references never seen in training
@article{ji2027failuremimic,
author = {Ji, Ziteng and Hong, Chuye and Shao, Yiyang and Sreenath, Koushil},
title = {{FailureMimic}: Humanoid Motion Tracking under Actuator Failure via Part-Wise Latent Residuals},
journal = {In-submission},
year = {2027},
}