Robotics papers — 2026-10-13

The Morning Preprint — this day as one edition: the day's papers, the five things worth knowing, and the episodes behind them.

Today's work focuses primarily on making robotic skills more flexible by tackling how to transfer knowledge from one object to a completely new one. This is important because if a robot policy only learns to handle one specific item, it cannot perform any other manipulation without retraining everything from scratch.

The main problem is that current vision-language action models often combine what something is, its meaning, and the actual steps needed into one unclear structure. This makes them good at tasks they have seen before but fragile when faced with small changes in the environment or object type.

To solve this decoupling problem, researchers are exploring w squared VLA which restructures how information moves through the model. Instead of letting all visual and language data feed into one large expert, this approach modulates the robot's state sequence using visual, spatial, and skill information in a more modular and understandable way. This method demonstrates that separating these knowledge representations allows for much more robust behavior cloning and enables zero-shot skill transfer to objects never seen during training.

This idea of learning strong foundation representations is also important because it aims to make models generalize across different sensors and environments. For example, DeFM is a self-supervised foundation model trained entirely on depth images to learn geometric and semantic representations that perform well in both simulation and the real world.

Finally, researchers are looking at how to catch when things go wrong in real-world applications where robots are actually deployed. FIDeL introduces a policy-independent module that uses anomaly detection and vision-language models to distinguish between normal deviations and actual failures during imitation learning. This shows significant improvements over existing failure detection methods on the BotFails dataset.

Today's papers

Decoupling the Declarative from the Procedural in Vision-Language-Action Models This paper introduces a new VLA model that separates what an agent knows about objects from how it performs actions, allowing for better generalization. [paper] [episode]
DeFM: Learning Foundation Representations from Depth for Robotics This work presents a foundation model trained only on depth images to learn robust representations for robotic tasks across various environments. [paper] [episode]
Failure Identification in Imitation Learning Via Statistical and Semantic Filtering This paper proposes a method to detect real-world failures in imitation learning by using anomaly detection and vision-language models to distinguish true failures from normal variations. [paper] [episode]

The papers

Decoupling the Declarative from the Procedural in Vision-Language-Action Models — The gist: w2VLA, a new Vision-Language-Action (VLA) model, successfully decouples declarative knowledge from procedural knowledge to enable zero-shot skill transfer to novel objects. [episode]
DeFM: Learning Foundation Representations from Depth for Robotics — The gist: DeFM, a self-supervised foundation model trained entirely on depth images, learns robust geometric and semantic representations that generalize across diverse robotics tasks without task-specific fine-tuning. [episode]
Failure Identification in Imitation Learning Via Statistical and Semantic Filtering — The gist The introduction proposes FIDeL, a policy-independent failure detection module that uses vision-based anomaly detection combined with semantic filtering to distinguish between benign deviations and genuine failures in imitation learning policies. How it works 1. [episode]

Important terms

Knowledge Transfer: This is about teaching a robot to use skills learned for one object and apply them successfully to a completely new, different object without needing extensive retraining.
Vision-Language Action Models (VLA): These are AI models that try to combine what an object looks like, what it means, and the actions needed into one confusing structure. They struggle when things change slightly.
$w^2$ VLA: This is a new way to structure how information flows in a robot's model. It separates visual, spatial, and skill data so the robot can handle different objects much more robustly.
Foundation Models (e.g., DeFM): These are large models trained on massive amounts of data to learn strong, general representations of things like geometry and meaning. They help robots work well in new environments.
Anomaly Detection in Imitation Learning: This technique is used to spot when a robot's actions are going wrong during real-world tasks. It helps distinguish between normal errors and actual system failures.