AI Paper Atlas

Research map / Reinforcement learning / Imitation learning

Revisiting Distribution Correction Estimation for Offline Imitation Learning with Suboptimal Dataset

Quang Anh Pham, Tien Anh Mai, Akshat Kumar

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

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add the mixture distribution (vs removing it) (ablation)ReDICE · D4RL offline IL tasksaverage return: removing it causes the largest drop (about 20 times)Improved
add the Gumbel regression loss (vs removing it) (ablation)ReDICE · D4RL offline IL tasksaverage returnImproved
add tau-scaled policy extraction (vs removing it) (ablation)ReDICE · D4RL offline IL tasksaverage returnImproved
raise tau up to 0.9 (ablation)ReDICE · D4RL offline IL tasksaverage return: peaks at tau = 0.9Improved
raise tau beyond 0.9 toward 1.0 (ablation)ReDICE · D4RL offline IL tasksaverage return: diminishing or negative returns after 0.9Hurt
set mixture coefficient beta to 0.5 rather than 0.1 or 0.9 (ablation)ReDICE · D4RL offline IL tasksaverage return: beta = 0.5 best vs 0.1 and 0.9Improved
larger beta (0.7 or 0.9) on low-coverage data (ablation)ReDICE · random+few-expert vs medium+few-expertaverage returnMixed
add tau-scaled policy extractionReCOIL · D4RL offline IL tasksaverage return: 67.76 -> 81.49Improved
replace ReCOIL objective (with tau-scaling) by the ReDICE objectiveReCOIL with tau-scaling · D4RL offline IL tasksaverage return: 81.49 -> 97.68Improved
use ReDICE instead of ReCOILoffline IL, 1 expert trajectory · D4RL suboptimal datasets averagenormalized score: 67.76 -> 97.68Improved
use ReDICE instead of ReCOIL under limited expert coverageoffline IL, no expert rollouts · D4RL limited coverage averagenormalized score: 28.16 -> 38.45Improved
use ReDICE instead of baselines on maritime navigationShipNaviSim · maritime navigationgoal reach rate: ≈60% -> ≈85%Improved

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