Out-of-Distribution Detection
Both families share limitations for vision-language-action models (VLAs): density-based detectors memorize the training-task distribution, while supervised detectors require out-of-distribution labels and transfer poorly across perturbations. We instead use task progress, a feature intrinsic to the policy, as a calibrated out-of-distribution signal requiring no out-of-distribution supervision and tying failure detection to the model’s own representation of task completion.
The Progress Feature
Task progress is a useful signal for runtime monitoring because it satisfies several desirable properties: it decreases monotonically along expert trajectories, yielding predictable in-distribution behavior; can be easily labeled from offline data; readily flags out-of-distribution states when progress stalls; and generalizes across tasks.
Out-of-Distribution Detection via the Progress Feature
We find the progress signal to be a strong indicator of when the policy goes out of distribution, enabling us to develop a method to flag out-of-distribution behavior based on the progress feature decoded via our probes.
The progress-deviation measure compares predicted progress with actual progress.
The rollout is flagged as out of distribution the first time the progress deviation exceeds the selected detection threshold.
While simple, this procedure can detect stalled progress using the probe alone. It is task-agnostic, assuming the detection threshold is known or easily estimated, and does not require out-of-distribution labels.
Discussion and Limitations
Task progress is linearly readable from the policy’s activations, emerges from large-scale robotic pre-training, supports an unsupervised out-of-distribution detector, and varies under language counterfactuals. The feature is not steerable, exposing a gap between what the policy represents and what it acts on.
Our work has several natural extensions, including evaluating progress decodability across vision-language-action backbones, validating the out-of-distribution detector on real-robot rollouts, and probing for richer task-conditional signals beyond scalar progress. Our work suggests linearly readable semantic features can be a lightweight handle for monitoring deployed visuomotor policies, complementing the heavier interpretability toolkit increasingly applied to vision-language-action models.
Models and Dataset
We run all experiments on VLABench. The simulator provides a primitive dataset of 10 primary tasks:
- Add Condiment: Pick up the target condiment and pour it over the dish.
- Select Fruit: Pick up the target fruit and place it on the plate.
- Select Drink: Pick up the target drink from the fridge.
- Insert Flower: Pick up the target flower and place it in the vase.
- Select Poker: Pick up the target playing card from the table.
- Select Mahjong: Pick up the target mahjong piece from the table.
- Select Book: Pull the target book from the shelf.
- Select Chemistry Tube: Touch the target tube in the chemical holder.
- Select Toy: Pick up the target toy and place it in the box.
- Select Painting: Press the button for the target painting.
The dataset contains 5,000 trajectories, 500 per task, with prompts and objects that vary across scenes to keep it diverse.
Decodability and Steerability Versus Observability and Controllability
Feature observability is defined as the existence of some map that takes an embedding and outputs the corresponding feature.
Weak Decodability
In contrast, we measure decodability with linear maps alone. This requires the feature to be linearly recoverable from the representation and gives a more stringent criterion than observability. Because we aim to minimize risk, we also relax the equality in observability: rather than reproduce the feature exactly, the map need only recover it with low error, so we target high correlation instead of equality.
Features may not be exactly recoverable from the representation, but as long as they are linearly decodable with low error, they are still useful for interpretation and control.
Feature controllability is defined as the ability of a map to propagate a feature through the activations by injecting a modified activation.
We move away from this definition because it only checks whether the feature can propagate; it does not:
- Judge the quality of the feature given its inputs.
- Require the feature to cause changes in the model’s behavior.
Strong Decodability
We call a feature strongly decodable if perturbing its inputs drives its output to the ground truth those inputs imply. This is a stronger condition because it lets the interpreter define how the feature should depend on its inputs, for example requiring task progress to track the instruction and the trajectory time, which yields more natural conditions on the feature.
Steerability
A feature is steerable if perturbing it produces a new feature value that elicits the same response the model would give under the matching ground-truth inputs. This establishes a form of causality between the feature and the model’s behavior, since perturbing the feature reproduces the response we would see if that value were the true feature.
Together, these variations impose stronger and more varied conditions for controlling a feature by probing both its robustness and its causality, which are crucial for interpretability and control.
Frequently Asked Questions
What makes task progress useful for runtime monitoring? It decreases monotonically along expert trajectories, can be labeled from offline data, flags stalled progress, and generalizes across tasks.
How does the progress feature support out-of-distribution detection? A rollout is flagged when the deviation between predicted and actual progress first exceeds the selected threshold.
What tasks are included in the VLABench dataset? The dataset contains 10 tasks involving condiments, fruit, drinks, flowers, playing cards, mahjong pieces, books, chemistry tubes, toys, and paintings.
How does strong decodability differ from weak decodability? Weak decodability evaluates linear recovery with low error, while strong decodability requires perturbing inputs to drive the output toward the implied ground truth.