HUMANOID CONTROL · ADVERSARIAL ROBUSTNESS

Still walking.
Off command.

Command-referenced observation attacks
for humanoid locomotion.

Small changes to what a robot senses can redirect its motion.
The controller stays frozen. The execution command stays unchanged.

Watch the comparisons Paper temporarily unavailable
Schematic overview of CoRA: command-following locomotion, a bounded attack on proprioceptive observations, and redirected stable motion under the same execution command.

Tap the overview to enlarge

Same execution command. Altered motion. Maintained stability.
Schematic illustration of behavior hijacking. Figure 1 in the manuscript. Explore the overview ↗

Unitree G1 · Simulation · Bounded proprioceptive perturbations

THE BEHAVIOR

Stable locomotion.
A different trajectory.

Remaining upright is only part of reliable control. CoRA studies whether a humanoid can sustain a specified deviation from its trusted motion command.

In these experiments, the original command requests forward motion. The attack induces turning, lateral movement, or both. The framework describes general original and target command pairs.

Watch the behavior.

Two targets. The same forward command.
Protected execution commandAttack target & toleranceActual measured behavior
Yaw target · Clean, Targeted PGD, Reference PPO, Reward-min, Target PPO, and CoRA-Match at matched timestamps.

State & duration. ON/OFF TARGET labels show the current state. Green segments in the time bar mark on-target intervals.

Motion & path. Cyan readouts track actual behavior against the purple target band. The inset traces the bird’s-eye path.

Fall termination. Reward-min panels retain the FALL / TERMINATED label after the episode ends.

Individual 20-second simulation rollouts, synchronized across methods at original speed. The close-up view adds actual velocity–target curves. Aggregate occupancy and dwell across runs appear in the results below.

Reference PPO vs. CoRA: follow the same episode through time

Matched frames at 3, 10, and 18 seconds connect the stepping posture to measured behavior and target status.

Yaw task: Reference PPO and CoRA-Match at 3, 10, and 18 seconds, with measured yaw rate, target interval, status, and path.
Yaw target · Reference PPO above, CoRA-Match below. Tap to enlarge.
Explore the recorded frames and bird’s-eye trajectories
Recorded Unitree G1 simulation frames showing CoRA-Match yaw motion in the top row and CoRA-Progress lateral motion in the bottom row.

Swipe to explore frames · Tap to enlarge ↔

Recorded simulation frames at matched timestamps: CoRA-Match on yaw and CoRA-Progress on lateral motion. Figure 6 in the manuscript.
Bird’s-eye trajectories comparing Clean, Targeted PGD, Target PPO, and CoRA on yaw and lateral targets.
Bird’s-eye trajectories from 20-second episodes. Paths share an initial position and heading. Figure 5 in the manuscript.

THE METHOD

A reference from
the controller itself.

Ask how the frozen controller would respond to the target command. Use that response to guide an observation attacker.

CoRA training framework: a training-only command-reference branch guides the bounded observation attacker, while simulated motion provides behavior feedback and the reference weight decays to zero.
A target command is substituted into a copied observation to obtain a reference from the same frozen controller. The attack changes proprioception while preserving the execution command and previous-action input. Reference guidance fades during training, while behavior feedback remains active. Robot poses and action-space geometry are schematic. Figure 2 in the manuscript. Open full-size diagram ↗

01 / REFERENCE

Query the desired response.

The controller provides a target-command action at the current state. CoRA-Match scores absolute agreement. CoRA-Progress scores how much of the nominal-to-reference gap is closed.

02 / LEARN

Reward the actual motion.

A learned attacker proposes observation residuals under magnitude and rate limits. Closed-loop feedback measures target motion, while regularization penalizes unstable or excessive perturbations.

03 / ANNEAL

Let the reference fade.

The reference weight decreases to zero. The final training phase optimizes the behavioral reward and regularization. Deployment uses the trained attacker without reference queries.

THE EVIDENCE

More time on target.

Compare sustained motion across three target settings.

Mean target occupancy

Share of evaluation steps satisfying motion and stability conditions.

mean target occupancy

s

mean longest uninterrupted target motion

All methods and run-to-run variation

Swipe left to see all columns ↔

MethodOccupancy (%) ↑Dwell (s) ↑Falls (%) ↓

Mean ± standard deviation over five runs. Each learned method uses five independently trained attackers and 100 evaluation episodes per run. Episodes last 20 s, with a 1 s warm-up. See Table 3 and the evaluation protocol for additional metrics.

WHAT THE ABLATIONS SHOW

Guide early.
Optimize behavior throughout.

The reference supplies an action-level learning signal. Motion feedback determines whether the intended behavior is realized.

Removing motion feedback sharply reduces occupancy. Keeping the reference weight constant also reduces occupancy on both tested targets.

Component ablation · target occupancy (%) · mean ± standard deviation over three runs
Training objectiveYawLateral
Target only86.6 ± 1.885.5 ± 2.4
Full CoRA89.3 ± 0.189.9 ± 1.1
No annealing84.2 ± 2.583.1 ± 3.7
Reference only2.3 ± 3.761.2 ± 5.5

Separate ablation runs from the main comparison. Source: Table 4.

Explore the episode-level distributions
Yaw task: episode-level occupancy, dwell, and target-acquisition distributions.
Yaw distributions · Tap to enlarge
Lateral task: episode-level occupancy, dwell, and target-acquisition distributions.
Lateral distributions · Tap to enlarge

THE TAKEAWAY

Command adherence.
Alongside stability.

Continued locomotion can coexist with sustained behavioral deviation. Command adherence adds a complementary measure of reliability for learned robot controllers.

CoRA

Command-Referenced Observation Attacks for
Behavior Hijacking in Humanoid Locomotion

Zepeng Wang1,* · Ziluo Ding2,* · Fengyi Zhang3 · Kan Yuan4 · Xiaochuan Shi1,†

1 Wuhan University
2 Beijing Innovation Center of Humanoid Robotics
3 Wuhan University of Technology
4 Shenzhen Polytechnic University

* Equal contribution · † Corresponding author

Paper temporarily unavailableRepository (restricted access) ↗
Research overview

CoRA learns bounded proprioceptive perturbations to induce specified humanoid motions while maintaining stable locomotion and leaving the execution command unchanged. A frozen controller supplies a target-command action reference during training. This guidance complements a reward based on measured motion and gradually decreases to zero. Comparative experiments on a simulated Unitree G1 controller show sustained behavioral deviation under bounded observation perturbations. CoRA-Match and CoRA-Progress achieve the highest mean target occupancy and longest uninterrupted target motion on the yaw and lateral tasks, respectively, among the evaluated methods.

Citation

@misc{wang2026cora,
  title = {Command-Referenced Observation Attacks for Behavior
           Hijacking in Humanoid Locomotion},
  author = {Wang, Zepeng and Ding, Ziluo and Zhang, Fengyi
            and Yuan, Kan and Shi, Xiaochuan},
  year = {2026},
  url = {https://xiaopeng-whu.github.io/CoRA/}
}