TL;DR
RECLAIM is a training-free, inference-time defense that protects robot Vision-Language-Action (VLA) models from adversarial patches by using the model's own internal attention as a feedback signal. By combining known robot geometry with local visual interventions, it recovers task performance across various attacks—raising success rates from 0% to 70% on challenging benchmarks—without requiring auxiliary models or expensive retraining.
What Problem Does This Paper Address?
Vision-Language-Action (VLA) models are the "brains" of modern embodied agents, translating visual scenes and text instructions into physical robot movements. However, these models are surprisingly fragile. A small, strategically placed "adversarial patch" in the camera's view can hijack the model's perception, causing "action drift" where the robot performs the wrong movement.
In a closed-loop system, this is catastrophic: one wrong move changes the robot's perspective, leading to a chain reaction of errors that ultimately fails the task. Existing defenses typically require retraining the VLA (adversarial training) or running heavy auxiliary models (like segmentors or diffusion models) at every control step to "clean" the image. These approaches are computationally expensive and difficult to deploy in real-time robotic systems with limited resources.
Methods
The authors introduce RECLAIM (Robot-Embodied Control Loop Attention-Informed Mitigation), a defense that treats the frozen VLA model itself as a diagnostic tool. It consists of three main components:
- Geometry-Guided Attention Localization (GAL): Instead of using an external object detector, RECLAIM looks at the VLA's own attention maps. Since robots naturally attend to their own grippers and the objects they are touching, GAL uses the robot's known physical geometry to "discount" this expected attention. Any remaining "residual" attention spikes are flagged as potential adversarial patches.
- Attention-Response-Guided Purification (ARP): Once a suspicious region is identified, RECLAIM tests several local "treatments" (e.g., blurring, graying out, or inpainting the area). It selects the treatment that most effectively reduces the suspicious attention while maintaining high confidence and consistency in the model's predicted actions.
- Memory-Assisted Temporal Revalidation (MTR): Because the robot and objects move, a patch's relative position changes. MTR maintains a memory of successful past interventions and re-validates them against new proposals, ensuring the defense remains stable and effective throughout the entire task rollout.
Datasets
The researchers evaluated RECLAIM using the LIBERO benchmark, specifically across four suites: Spatial, Object, Goal, and Long. These suites cover 40 distinct manipulation tasks ranging from simple picking to long-horizon multi-step interactions.
To test robustness, they subjected the models to four different adversarial attack methods:
- UADA (Untargeted Action Discrepancy Attack)
- EDPA (Embedding Disruption Patch Attack)
- RFAS (Universal patch transfer attack)
- UPA (Untargeted Patch Attack)
Key Insights and Ideas
- The Policy as its Own Defender: The core "aha moment" is that a VLA's internal attention response to local visual edits provides enough feedback to guide its own purification. You don't need a second model to tell you if an image is "clean"; you just need to see if the VLA stops being distracted.
- Embodied Context Matters: By incorporating robot geometry (joint positions and camera projection), the system can distinguish between "good" attention (on the gripper) and "bad" attention (on a malicious patch).
- Efficiency through Sparsity: Because RECLAIM only intervenes in small local regions and reuses successful treatments over time, it is significantly faster than methods that process the entire image every frame.
Key Results
- Dramatic Recovery: On the LIBERO-Object suite, RECLAIM raised the rollout success rate from 0% (undefended) to 70.0% under UADA attacks.
- High Task Coverage: RECLAIM achieved at least one successful rollout on 95% of all evaluated tasks, compared to 0% for the undefended baseline.
- Speed and Efficiency: RECLAIM is 77.31% faster than the previous state-of-the-art defense (PAD), with a mean decision time of only 548.9 ms per control step. Crucially, it requires zero additional GPU memory for auxiliary models.
- Generalization: The defense proved effective across different VLA architectures, including OpenVLA and NORA, demonstrating that the underlying principle of attention-guided feedback is broadly applicable.
Limitations and Future Work
- Long-Horizon Challenges: While RECLAIM performs exceptionally well on most tasks, its success rate is lower on "Long" horizon tasks (38.33%). The authors note that in these tasks, even small residual errors in purification can compound over hundreds of control steps.
- Dependence on Attention Quality: The defense relies on the VLA having somewhat interpretable and "useful" internal attention. If a future model's attention maps are completely diffuse or uninformative, GAL might struggle to localize patches.
- Future Directions: The authors suggest that future work could explore using this embodied feedback to not just mask patches, but to actively move the robot's camera to better viewpoints that naturally mitigate the attack.
Summary updated