Skip to main content
Back to timeline
arXivSource publication:

A frozen robot policy becomes its own defender: RECLAIM lifts LIBERO-Object success from 0% to 70.0% without retraining

Synopsis

An adversarial patch placed in a robot's camera view can amplify a wrong action into a wrong next observation through the perception-action loop. RECLAIM leaves the frozen vision-language-action policy's parameters untouched and trains no extra model; it uses the policy's own attention shifts and action uncertainty under local image edits to decide where a patch is and which local repair to apply, then rechecks those choices as motion changes the scene. On OpenVLA across four patch attacks and the LIBERO task suites, it succeeds in at least one rollout on 95% of evaluated tasks and raises LIBERO-Object success from 0% without defense to 70.0%, at a mean decision time of 548.9 ms per control step.

Source-provided article image: RECLAIM: Robot-Vision-Informed Adversarial Patch Defense via Embodied Attention Feedback
arXiv

Interpretation

In robot manipulation an adversarial patch does more than corrupt a single frame: it changes the robot's action, which changes the next observation, so the initial disturbance compounds. RECLAIM puts the defense on the input side, editing the observation before it reaches the frozen vision-language-action policy so the policy resumes following the instruction. Earlier defenses either fine-tune the visual encoder with clean and perturbed training pairs, or depend on segmentation models, diffusion models or an external vision-language model for auxiliary inference, adding training or online computation. Under matched OpenVLA and UADA patch-attack conditions, rollout success across the four LIBERO suites is 0.00% without defense, 52.92% with RECLAIM and 40.83% with accelerated PAD; RECLAIM covers 95.00% of tasks versus PAD's 80.00%.

Attention strength alone is ambiguous for locating a patch, because the arm, gripper and manipulated object also attract attention. RECLAIM first subtracts the attention explained by known robot geometry, then looks for residual attention that persists on particular regions and treats those as candidates for inspection. For each candidate region it does not commit to one repair; it tries several local operators, including gray blending, blur, median filling and Telea inpainting, and compares which treatment suits the current view. The comparison uses whether attention mass on the edited region drops, whether action-prediction entropy falls, and whether different operator families leave the policy attending similarly outside the region; under UADA, a gray-only variant reaches 37.50% success, a Telea-only variant 47.92%, and full RECLAIM 52.92%.

Once the robot moves, a previously chosen region and operator may no longer fit, so the defense has to remember recent choices and rescore them. RECLAIM keeps a memory bank of recently selected region-operator pairs, compares old and new candidates on the current image at full decisions, and between full decisions keeps a treatment that still passes geometry and action-quality checks. Removing this temporal revalidation and rerunning localization and selection at every control step sharply reduces both success and mask alignment. Without temporal revalidation, success falls from 52.92% to 4.17%, mask IoU from 50.83% to 16.90%, and policy queries per control step rise from 2.107 to 9.572; full RECLAIM averages 548.90 ms per control step versus 2,419.64 ms for accelerated PAD.

Perspective

The evaluation covers a frozen OpenVLA on 40 tasks across the four LIBERO suites, plus a NORA comparison under the same UADA setting, with patch positions fixed in image coordinates before evaluation. The intended readers are researchers and engineers concerned with embodied-policy deployment security and inference-time intervention.

On clean inputs the Long suite's success drops from 50.00% to 30.00%, which the authors suggest may reflect longer action sequences amplifying small perception or purification errors; NORA reaches only 5.00% clean success on Long, so its S/O/G subset is treated as the more informative assessment of defense performance.