analiza aiqbench
ConTraIRL addresses the unreliability of reward transfer in Inverse Reinforcement Learning by decoupling latent representations of environment dynamics and task goals. This dual-encoder architecture enables compositional reward inference, allowing policies to generalize effectively to unseen combinations of dynamics and objectives with improved sample efficiency. The approach offers a practical solution for few-shot transfer scenarios in continuous control benchmarks where standard transfer methods fail.
ConTraIRL: Factorized Contrastive Abstractions for Transferable IRL
arXiv:2606.03017v1 Announce Type: cross Abstract: Reward transfer in Inverse Reinforcement Learning (IRL) is unreliable when policies must generalize to unseen combinations of environment dynamics and task goals. We propose Factorized Contrastive Abstractions for Transferable IRL (ConTraIRL), a framework that enables compositional reward transfer by learning decoupled latent representations of these two factors. ConTraIRL uses a dual-encoder architecture that maps observations into separate dynamics and goal latent spaces, trained with a dual contrastive objective. Temporal alignment encourages the dynamics encoder to learn goal-invariant structure, while the goal encoder captures dynamics-invariant features. This factorization supports reward inference under recombined dynamics-goal settings. Experiments on continuous control benchmarks demonstrate effective few-shot transfer to unseen dynamics-goal pairings, improving sample efficiency and reward recovery over transfer IRL baselines.
Abstract:Reward transfer in Inverse Reinforcement Learning (IRL) is unreliable when policies must generalize to unseen combinations of environment dynamics and task goals. We propose Factorized Contrastive Abstractions for Transferable IRL (ConTraIRL), a framework that enables compositional reward transfer by learning decoupled latent representations of these two factors. ConTraIRL uses a dual-encoder architecture that maps observations into separate dynamics and goal latent spaces, trained with a dual contrastive objective. Temporal alignment encourages the dynamics encoder to learn goal-invariant structure, while the goal encoder captures dynamics-invariant features. This factorization supports reward inference under recombined dynamics-goal settings. Experiments on continuous control benchmarks demonstrate effective few-shot transfer to unseen dynamics-goal pairings, improving sample efficiency and reward recovery over transfer IRL baselines.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Robotics (cs.RO)
arXiv:2606.03017 [cs.LG]
(or arXiv:2606.03017v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2606.03017 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 216 cuvinte
Subiecte conexe
ConTraIRL: Factorized Contrastive Abstractions for Transferable IRL
arXiv · 30 sept. 2026
Game Design Creativity evals
LessWrong · 12 oct. 2026
Learning Cross-Model Activation Alignments with Explicit Many-to-Many Layer Maps
arXiv · 8 oct. 2026
Probabilistic Contrastive Pretraining for Multi-task ADME Property Prediction
arXiv · 12 iun. 2026