News
- Prof. Han-Lim Choi’s Research Team, Introduces PAMD for Visual Reinforcement Learning Paper Accepted to ICML 2026
- 관리자 |
- 2026-06-17 14:14:32|
- 32
- 2026-06-17 14:14:32|
Daegyeong Roh, Juho Bae, and Prof. Han-Lim Choi from our department’s Laboratory for Information and Control Systems present “PAMD: Structured Adaptive Distances for Bisimulation Representations in Visual Reinforcement Learning,” accepted to ICML 2026. The work targets a core bottleneck in pixel-based reinforcement learning: learning latent states whose distances reflect behavioral similarity.
Instead of relying on fixed ℓp norms or unconstrained pairwise regressors, PAMD learns a pair-conditioned positive-definite Mahalanobis geometry. This gives the representation an adaptive but structured notion of distance, expressive enough to capture context-dependent anisotropy while preserving learning pressure on the encoder.
As a plug-in module for DBC, MICo, and SimSR-style objectives, PAMD improves performance across image-based DeepMind Control Suite tasks. It reports 878±8 on Cheetah Run and 250±3 on Hopper Hop, showing competitive performance against strong visual RL baselines.
Instead of relying on fixed ℓp norms or unconstrained pairwise regressors, PAMD learns a pair-conditioned positive-definite Mahalanobis geometry. This gives the representation an adaptive but structured notion of distance, expressive enough to capture context-dependent anisotropy while preserving learning pressure on the encoder.
As a plug-in module for DBC, MICo, and SimSR-style objectives, PAMD improves performance across image-based DeepMind Control Suite tasks. It reports 878±8 on Cheetah Run and 250±3 on Hopper Hop, showing competitive performance against strong visual RL baselines.