Why VLA Safety Monitors Break Under Task Shift: Testing a Modality-Imbalance Explanation

Robot policies built on large vision-language models can be inspected from the inside. A simple probe on their hidden activations can flag when the arm is about to collide, push too hard, or leave its workspace, and the probe can then trigger a stop. In OpenVLA, a vision-language-action model (a language model that reads camera images and an instruction and emits motor commands as tokens), such internal monitors work well on familiar tasks. They stop working when the robot faces a new family of tasks. We propose a specific, testable cause. Inside these models, language activations are hundreds of times larger in magnitude than visual activations. A monitor fit on the raw hidden state therefore mostly learns which instruction it is looking at, and only weakly learns what the scene looks like. When the instructions change, the monitor has nothing left to go on.

If this is right, the fix is cheap and structural. Give each modality equal weight before learning the monitor, and read risk from the interaction between the instruction and the scene rather than from the instruction alone. A monitor built this way should transfer across task families. It should also catch hazards that come from a combination of inputs, where an ordinary scene and an ordinary command together produce a dangerous motion. The project measures the imbalance, tests its causal role with targeted interventions, and checks whether the corrected monitor reduces physical violations on LIBERO, SimplerEnv and SafeLab.

Core Claim

Sparse Progress and Risk Features in OpenVLA trains sparse autoencoders on intermediate OpenVLA activations. A sparse autoencoder is a model that rewrites activations as a small set of interpretable features. The paper finds features that track task progress and physical risk. A linear readout of progress reaches 0.932 AUROC within a task family. A risk-gated monitor cuts simulated violations by 16% while keeping 92% of baseline success. The progress readout, however, falls to 0.537 AUROC when training uses LIBERO Goal/Object and testing uses Spatial/Long. That paper does not explain the collapse. The file reports no family-shift number for the risk features specifically, so measuring it is our first step.

A separate study, Modality-Aware Block Rotation for VLA Quantization, shows that language-channel activations in early OpenVLA layers exceed visual ones by more than 600×. It also shows that mixing the two modalities with a global rotation contaminates the visual channels. That paper studies quantization accuracy, not safety.

Our claim joins these two findings. The family-shift collapse is largely caused by language energy dominating the dictionaries and probes that monitors are built on. Normalizing each modality separately, and reading risk from cross-modal interaction terms, should restore transfer. The novelty is the causal account and the repair, both evaluated on physical outcomes. We do not claim a new detector architecture.

There is a second reason this matters. Nouns, Not Modifiers shows that OpenVLA matches nouns to objects but drops spatial modifiers. Accuracy is 15% on "left" instructions versus 80% on "right", and 93% of failures reach for the wrong object of the correct category. In a cluttered or fragile scene, reaching for the wrong same-category object is itself a physical hazard. A monitor that ignores the visual side cannot see it.

Research Program

Measuring the imbalance. For each OpenVLA layer, we partition hidden-state channels and token positions by source: language prompt, visual patches, and generated action tokens. We report each source's share of activation energy and of sparse-autoencoder reconstruction loss. We then attribute every risk and progress feature to its dominant source.

The prediction is concrete. Features whose readouts collapse under family shift should be language-dominated. Features that survive should carry visual or action-token mass.

Causal test. We retrain the sparse autoencoders and linear probes in three settings, holding everything else fixed:

  • Language down-weighted. Language channels are rescaled to unit per-modality variance using statistics from the training families only.
  • Language removed. The language block is removed entirely.
  • Scrambled control. The same rescaling is applied to a randomly chosen channel block of equal size.

If modality imbalance is the cause, only the first two settings should recover family-held AUROC. The scrambled control should not.

Interaction monitor. We build a monitor from a few signals:

  • Inputs. Per-modality whitened features, plus the attention mass that action tokens place on visual patches. The second signal reflects whether the instruction actually selected a region of the scene.
  • Readout. A sparse linear readout combines these features with their pairwise products.

This targets hazards that depend jointly on instruction and scene. The idea is the embodied counterpart of the "safe inputs, unsafe output" failures documented in text form by Cross-Modality Safety Alignment.

Closed-loop use. At deployment, the monitor gates a bounded fallback that stops the arm and retracts it slightly. Thresholds are calibrated on held-in families only, so no information leaks from the test families.

Baselines

Raw-feature monitor. The reference method is the sparse-autoencoder risk monitor from Sparse Progress and Risk Features in OpenVLA, reimplemented on the same layer with the same dictionary size and training data. It is the strongest competitor because it is the only internal physical-risk monitor for vision-language-action models in the file. Its within-family numbers are 0.932 AUROC for progress, and 16% fewer violations at 92% retained success. We give it the same hyperparameter-search budget as our monitor. To beat it, our monitor must exceed its family-held risk AUROC by a clear margin across three seeds, and must lower violations at matched retained success on the held-out families.

Naive references. The "do nothing" reference is the unmonitored OpenVLA policy fine-tuned on LIBERO. That paper reports 76.5% average success in FP16. A second naive monitor thresholds end-effector velocity and proprioception. That signal reached 0.711 AUROC for progress in the same paper; no reported risk number was found on Lacuna for it.

Gap-closing fine-tune. We fine-tune the projector of OpenVLA with an L2 penalty on the image-text gap, as in REGAP. REGAP cut unsafe rates by 16.3% on HADES Toxic for LLaVA-v1.5-7B, but in a text-only setting. We then retrain the raw monitor on top of the fine-tuned model. This tests whether shrinking the gap in the model, rather than rebalancing inside the monitor, gives the same transfer.

Ablations. Ablations isolate the key ingredients: per-modality normalization without interaction terms, interaction terms without normalization, and the scrambled-block control.

Why not just…?

Why not just put a Model Predictive Control or Hamilton-Jacobi reachability filter in front of the actuators? Filters such as the one in Safe and Performant Deployment via MPC and HJ Reachability enforce geometric limits well when the state is known. They cannot tell that the arm is heading toward the left can when the user asked for the right one, because that hazard is semantic. PACT also notes that such filters depend on privileged, noise-free object states. We run an oracle-state filter on SafeLab as a complementary upper reference and report which violation types each method catches. We expect the filter to dominate contact-force violations and to miss wrong-target events.

Why not just train the monitor on visual features alone, since language is the problem? This is the most serious threat, and it is a direct ablation in our study. If a vision-only probe matches the interaction monitor, the conclusion becomes "drop language" and the interaction component is unnecessary. We expect vision-only to fail on paired-instruction scenes, where the scene is identical and only the instruction makes the motion dangerous. The left/right pairs from Nouns, Not Modifiers and pour-versus-place pairs on the SafeLab glassware tasks are the test cases. If vision-only does not fail there, we report that result.

Why not just fine-tune the policy on safe demonstrations or with safety-penalized RL? The file gives three reasons for doubt:

  • Adversarial Attacks on Robotic VLA Models finds that domain fine-tuning increases vulnerability through action-space concentration.
  • The sparse-feature paper found no return gain from offline RL guided by its features.
  • On SafeLab, bounded residual RL does raise safe success by 33.1–43.0 points for Diffusion Policy and π0.5.

The RL result is a real competitor for the policy itself. A monitor, however, plays a different role: it is meant to detect what the policy gets wrong after deployment. We therefore also apply our monitor on top of a residual-RL-improved policy. The test is whether it still lowers residual violations under family shift.

Proposed Experiments

LIBERO. We use the family-held split from the sparse-feature paper: train on Goal/Object, test on Spatial/Long, with the task-held split as a control. We report risk and progress AUROC, violation rate, and success.

SimplerEnv. We use the matched left/right and distance-modifier scenes from Nouns, Not Modifiers and relabel wrong-target reaches as hazards. Each monitor's detection rate on these events is the key cross-modal test.

SafeLab. We fine-tune OpenVLA on the 6,400 expert trajectories of SafeLab and hold out manipulation categories as the analogue of task families. The metric is safe success rate, the share of episodes that succeed with no spill, breakage or force violation. SafeLab reports that for current policies safe success trails raw success by more than 30 points.

Decisive checks. The account stands if three things hold at once. Raw-feature risk AUROC must collapse under family shift. Collapsing features must be predominantly language-sourced. Per-modality normalization, but not the scrambled control, must recover most of the lost AUROC.

Falsifying outcome. If normalization recovers nothing, the imbalance account is wrong. The collapse then more likely reflects that the visual risk features themselves do not transfer, and the paper reports it as such.