A naukri.com initiative
Arxiv
2d
227
Image Credit: Arxiv
Self-Predictive Representations for Combinatorial Generalization in Behavioral Cloning
- Behavioral cloning (BC) with supervised learning is used to learn policies from human demonstrations in domains like robotics.
- Goal-conditioning BC policies enables capturing diverse behaviors within an offline dataset.
- While goal-conditioned behavior cloning methods perform well on in-distribution tasks, they may not generalize zero-shot to tasks requiring conditioning on novel state-goal pairs (combinatorial generalization).
- Temporal consistency in state representations learned by BC plays a role in enabling combinatorial generalization.
- Encouraging temporal consistency reduces the out-of-distribution gap for novel state-goal pairs.
- Successor representations that encode the distribution of future states from the current state help in achieving temporal consistency.
- Prior methods for learning successor representations have used contrastive samples, temporal-difference learning, or both.
- A new approach, BYOL-γ augmented GCBC, is proposed in this work for representation learning without contrastive samples or TD learning.
- BYOL-γ augmented GCBC can theoretically approximate the successor representation in the finite MDP case.
- Empirical results show competitive performance across various challenging tasks requiring combinatorial generalization with BYOL-γ augmented GCBC.
Read Full Article
13 Likes
For uninterrupted reading, download the app