Our new X account is live! Follow @wizwand_team for updates
WorkDL logo mark

LobsDICE: Offline Learning from Observation via Stationary Distribution Correction Estimation

About

We consider the problem of learning from observation (LfO), in which the agent aims to mimic the expert's behavior from the state-only demonstrations by experts. We additionally assume that the agent cannot interact with the environment but has access to the action-labeled transition data collected by some agents with unknown qualities. This offline setting for LfO is appealing in many real-world scenarios where the ground-truth expert actions are inaccessible and the arbitrary environment interactions are costly or risky. In this paper, we present LobsDICE, an offline LfO algorithm that learns to imitate the expert policy via optimization in the space of stationary distributions. Our algorithm solves a single convex minimization problem, which minimizes the divergence between the two state-transition distributions induced by the expert and the agent policy. Through an extensive set of offline LfO tasks, we show that LobsDICE outperforms strong baseline methods.

Geon-Hyeong Kim, Jongmin Lee, Youngsoo Jang, Hongseok Yang, Kee-Eung Kim• 2022

Related benchmarks

TaskDatasetResultRank
Stacking2CAUSALWORLD (in-distribution (Space A to Space A))
Average Return958.6
14
Manipulator Insert BallDeepMind Control Suite (in-distribution)
Average Return197.8
7
Manipulator Insert PegDeepMind Control Suite (in-distribution)
Average Return299.2
7
Cartpole SwingupDeepMind Control Suite (in-distribution)
Average Return293
7
Fish SwimDeepMind Control Suite (in-distribution)
Average Return188.8
7
Cheetah RunDeepMind Control Suite (in-distribution)
Average Return74.53
7
Finger Turn hardDeepMind Control Suite (in-distribution)
Average Return190.9
7
PickingCAUSALWORLD space B
Average Return337.8
7
TowersCAUSALWORLD (in-distribution (Space A to Space A))
Average Return735.8
7
Walker StandDeepMind Control Suite (in-distribution)
Average Return252.3
7
Showing 10 of 26 rows

Other info

Follow for update