CoRL 2025 Papers

In meta-learning and its downstream tasks, many methods use implicit adaptation to represent task-specific variations. However, implicit approaches hinder interpretability and make it difficult to understand which task factors drive performance. In this work, we introduce a disentangled multi-context meta-learning framework that explicitly learns separate context vectors for different aspects that define a task. By decoupling these factors, our approach improves both robustness, through deeper task understanding, and generalization, by enabling context vector sharing across tasks with the same context. We evaluate our approach in two domains. First, on a sinusoidal regression benchmark, our model outperforms baselines on out-of-distribution tasks and generalizes to unseen sine functions by sharing context vectors associated with shared amplitudes or phase shifts. Second, in a quadruped locomotion task, we disentangle the robot-specific properties and the characteristics of the terrain in the robot dynamics model. Using these context vectors in reinforcement learning, the learned policy demonstrates improved robustness under out-of-distribution conditions, compared to a model using a single unified context. Furthermore, by effectively sharing context, our model enables successful sim-to-real policy transfer to challenging terrains with out-of-distribution robot-specific properties using only real data from flat terrain, which is not achievable with single-task adaptation.

Skill effect models for long-horizon manipulation tasks are prone to failures in conditions not covered by training data distributions. Therefore, enabling robots to reason about and learn from failures is necessary. We investigate the problem of efficiently generating a dataset targeted to observed failures. After fine-tuning a skill effect model on this dataset, we evaluate the extent to which the model can recover from failures and minimize future failures. We propose Fail2Progress, an approach that leverages Stein variational inference to generate multiple simulation environments in parallel, enabling efficient data sample generation similar to observed failures. Our method is capable of handling several challenging mobile manipulation tasks, including transporting multiple objects, organizing a constrained shelf, and tabletop organization. Through large-scale simulation and real-world experiments, we demonstrate that our approach excels at learning from failures across different numbers of objects. Furthermore, we show that Fail2Progress outperforms several baselines.