AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

Masked Skill Token Training for Hierarchical Off-Dynamics Transfer

Zeyu Feng, Haiyan Yin, Yew-Soon Ong, Harold Soh

ICLR 2026 · Reinforcement learning › Visual and representation RL · approach: Skill learning and diffusion policies

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 masked skill tokens and feasibility-aware planning (MSTT) versus DiffuserMSTT · Maze2D (avg)return: 33.12 -> 140.35Improved
fine-tune Diffuser on state-only target demonstrationsDiffuser · Maze2D (avg)return: 33.12 -> 57.11Improved
MSTT versus state-only fine-tuned Diffuser on goal rateMSTT · Maze2D (avg)goal success rate: 72% -> 94.33%Improved
MSTT versus action-labelled BC-ta on goal rateMSTT · Maze2D (avg)goal success rate: 82% -> 94.33%Improved
MSTT versus fine-tuned Diffuser on Maze2D Target 1 goal rateMSTT · Maze2D Target 1goal success rate: 99% -> 89%Hurt
MSTT versus Diffuser on FetchReach goal rateMSTT · FetchReachgoal success rate: 99% -> 88%Hurt
MSTT versus Diffuser on FetchReach failure rateMSTT · FetchReachfailure rate: 61% -> 0%Improved
MSTT versus inverse dynamics on ReplicaCAD goal rateMSTT · Habitat ReplicaCADgoal success rate: 23% -> 75%Improved
increase VQ-VAE codebook size from 15 to 240 (ablation)MSTT · Maze2D Target 2goal success rate: 0% -> 95%Improved
increase VQ-VAE codebook size from 240 to 1000 (ablation)MSTT · Maze2D Target 3goal success rate: 99% -> 86%Hurt
increase VQ-VAE codebook size from 240 to 512 (ablation)MSTT · Maze2D Target 2goal success rate: 95% -> 90%No clear effect
use VLM-generated demonstration instead of provided demonstrationMSTT · Maze T1goal success rate: 89% -> 94%No clear effect

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