ExecGap attacks VLA robots on the execution side: patches push commands that look valid but the body cannot realize
Synopsis
ExecGap moves an adversarial patch's target from prediction to execution: using kinematically inefficient directions and dynamically weakly actuated directions of the robot, it biases the task-space commands issued by vision-language-action models toward motions the embodiment cannot realize effectively. Across the four LIBERO suites in simulation it produces larger command-execution mismatch and weaker realized execution than existing patch attacks on OpenVLA, OpenVLA-OFT and a flow-based policy; on Kinova Jaco2 and ROKAE xMate ER7 Pro, a printed patch without real-world re-optimization cuts task success from 64% to 28%, from 74% to 36%, and from 86% to 48%.
Interpretation
An adversarial patch no longer has to make the model predict wrongly; it can make the model's command hard for the robot to carry out, weakening task progress. Earlier VLA patch attacks mainly optimized prediction-level discrepancy, semantic confusion or rollout-level deviation, treating the model output or rollout behavior as the target and leaving the physical realization of the command largely unexplored. On the LIBERO Spatial, Object, Goal and Long suites, ExecGap attains the largest command-execution direction error and normalized magnitude mismatch on all four suites for OpenVLA and the lowest command-execution gain on every suite; because several baselines already drive task success near zero, the distinction shows up in these execution-side metrics.
The patch is optimized against a surrogate that avoids differentiating through downstream execution, pulling predicted commands toward low-executability directions estimated at the current robot state. Directly optimizing the command-execution mismatch would require differentiating through downstream execution and closed-loop rollout dynamics, so ExecGap instead uses a robot-aware surrogate built from kinematic inefficiency and dynamic weak actuation. The kinematic term takes the direction of the smallest singular value from a Jacobian decomposition, and the dynamic term Boltzmann-weights the normals of the supporting facets closest to the origin in the torque-limited task-space acceleration polytope; across 400 OpenVLA-OFT states with state and command magnitude fixed, low-executability directions give the largest mean direction error on all four suites.
Perspective
The demonstrated scope is simulation on the four 10-task LIBERO suites plus three real-robot pairings: Kinova Jaco2 with OpenVLA, and ROKAE xMate ER7 Pro with OpenVLA and with a flow-based policy, each with 50 clean and 50 attacked trials, with the printed patch placed at a fixed, visible, non-occluding tabletop location and no real-world re-optimization. The intended readers are researchers and engineers concerned with the adversarial robustness of robot manipulation policies.
ExecGap requires white-box access to the VLA policy and to robot state and model information to estimate low-executability directions; physical validation is limited to a set of robot platforms and tabletop scenarios; and the work focuses on exposing execution-degradation attacks rather than studying detection or defense.
