AI Paper Atlas

Research map / Reinforcement learning / Imitation learning

Reconciling Safety and Performance via Dual-Expert Offline Imitation Learning

Seokin Seo, Sung Kuk Shyn, Minji Seo, Yisoo Lee, Jongmin Lee, Hongseok Yang, Kee-Eung Kim

NeurIPS 2026 · Reinforcement learning › Imitation learning · approach: Offline imitation from imperfect data

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a dual-expert offline imitation learning method that reconciles safety and performance without explicit cost functions or jointly safe-and-high-performing expert demonstrations.

Problem

Offline imitation learning must achieve high returns while providing safety guarantees, but explicit cost functions are hard to design for complex safety constraints and demonstrations that are both safe and high-performing are costly to collect.

Approach

Uses two separate expert demonstration sources (dual experts) so that safety and performance can be learned from distinct data rather than requiring a single expert that is both safe and high-performing; avoids hand-designed cost functions.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace distribution matching to d_S with safe-support constraint (DexDICE vs DemoDICE-S) (ablation)DexDICE · Re-curated DSRL, Safety GymnasiumAvg return: 0.27 -> 0.37Improved
replace distribution matching to d_S with safe-support constraint (DexDICE vs DemoDICE-U) (ablation)DexDICE · Re-curated DSRL, Safety Gymnasium# Safe tasks: 10/21 -> 17/21Improved
use dual-expert DICE instead of pseudo-labelled safe RL (CDT/TraC)DexDICE · Re-curated DSRL, Safety GymnasiumAvg return: 0.19 -> 0.37Improved
use dual-expert DICE instead of pseudo-labelled safe RL (TraC)DexDICE · Re-curated DSRL, MetaDriveAvg return: 0.23 -> 0.50Improved
add support-filtering to BC (FBC) vs plain BC (ablation)BC · Re-curated DSRL, Safety GymnasiumAvg return: 0.46 -> 0.13Hurt
DexDICE vs DemoDICE-S on MetaDriveDexDICE · Re-curated DSRL, MetaDrive# Best among safe: 5/9 vs 5/9No clear effect
DexDICE on MetaDrive vs CDT/TraC costDexDICE · Re-curated DSRL, MetaDrive# Safe tasks: 9/9 (CDT, TraC) vs 7/9Hurt
reduce safe demonstrations to N_S=2 (ablation)DexDICE · Controlled Safety Gymnasium (6 tasks)safety / returnHurt
degrade quality of D_S or D_P demonstrations (ablation)DexDICE · Controlled Safety Gymnasium (6 tasks)safety and returnNo clear effect
real-robot teleoperation vs baselinesDexDICE · Real mobile-robot navigation (10 trials)goal reaching and violationsImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy