New technique constrains diffusion models to match training data, boosting robotic reliability

A team of researchers has introduced Curiosity-Diffuser, a method designed to make imitation-learning policies safer and more predictable by steering diffusion models away from generating out-of-distribution behaviors. The work addresses a critical vulnerability in neural network-based robotics: the tendency of learned policies to "hallucinate" novel actions that deviate from training examples, risking unsafe real-world outcomes.

The core problem#

Neural networks powering robotic control policies can generalize in unpredictable ways, creating behaviors unseen in training data. This overgeneralization is particularly risky when deploying systems in physical environments where errors can cause harm. The researchers argue that constraining generated trajectories to stay close to the training distribution is essential for safe deployment.

How it works#

The proposed method combines two key components:

• Curiosity measurement: A Random Network Distillation (RND) module assesses how "novel" or unexpected each generated trajectory is relative to the training data.
• Guided generation: The system uses classifier-guided diffusion to minimize curiosity scores during inference, effectively penalizing behaviors that stray from the learned distribution.

This approach directs the diffusion model to explore only behaviors that resemble those in the training set, reducing the risk of harmful hallucinations.

Evaluation and results#

The authors introduce a computationally efficient metric that quantifies policy reliability by measuring behavioral similarity to the training dataset. Simulations and real-world experiments demonstrated that Curiosity-Diffuser improves task performance while keeping generated behaviors closer to the original training distribution. The consistency between model predictions and human-guided imitation data suggests the method effectively mitigates overgeneralization risks.

Significance#

The work addresses a foundational challenge in safe robotics: how to maintain robustness without sacrificing performance. By grounding learned policies in the training data distribution, the approach offers a path toward more dependable autonomous systems suitable for safety-critical applications.