Research map / Reinforcement learning / Offline RL

DaCe-DT: Data-Centric Offline Multi-Task Reinforcement Learning via Adaptive Prompts and Trajectory Correction for Heterogeneous Tasks

NeurIPS 2026 · Reinforcement learning › Offline RL · approach: Offline RL sample complexity theory

Read the paperCheck your idea against 108,967 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 DaCe-DT, a data-centric offline multi-task RL framework on Decision Transformer that adds adaptive prompt length masking, embedding-based prompt retrieval, and value-based return calibration, evaluated on Meta-World MT5/MT30/MT50.

Problem

Offline multi-task RL performance is limited by data-level issues: fixed prompt lengths across heterogeneous tasks, semantically irrelevant prompt segments, and fragmented trajectories that hinder stitching.

Approach

Combines a learnable per-task prompt length gate (LGPM), GPT-2 embedding-based retrieval of return-filtered prompt segments (RAPC), and a task-conditioned value function with per-task adaptive expectiles to correct return-to-go at inference (VARC) on top of Prompt-DT.

Key results

• On MT30, DaCe-DT reaches 93.33% success on near-optimal and 75.67% on sub-optimal data, reported as 11.73% and 13.34% improvements over the prior state of the art.
• On MT50 (random goals), DaCe-DT achieves 88.40% near-optimal and 64.00% sub-optimal success versus 78.80% and 57.50% for the best HarmoDT variants.
• Ablation on MT30: LGPM alone 83.7, RAPC alone 85.3, VARC alone 82.6, all three 93.3 success rate.

Limitations

• Bounded by quality and diversity of pre-collected offline data; cannot explore the environment.
• Evaluation confined to simulation benchmarks with well-defined task boundaries and rewards; real-robot and open-world settings untested.

Experimental setup

Datasets: Meta-World offline datasets from SAC-Replay (near-optimal and sub-optimal) Maze2D custom layouts

Benchmarks: Meta-World MT5 Meta-World MT30 Meta-World MT50 Meta-World MT50-rand Cheetah-dir Cheetah-vel Ant-dir Maze2D

Metrics: average success rate episodic return

Baselines: MTBC MTIQL MTDIFF-P MTDT Prompt-DT HarmoDT-R HarmoDT-M HarmoDT-F CARE PaCo Soft-M D2R

Builds on: Decision Transformer Prompt-DT HarmoDT Implicit Q-Learning

Closest papers by meaning

• Generalizable Task Representation Learning for Offline Meta-Reinforcement Learning with Data Limitations (AAAI 2024 · similarity 0.89)
• Simple Ingredients for Offline Reinforcement Learning (ICML 2024 · similarity 0.89)
• Temporal Distance-aware Transition Augmentation for Offline Model-based Reinforcement Learning (ICML 2025 · similarity 0.88)
• Towards an Information Theoretic Framework of Context-Based Offline Meta-Reinforcement Learning (NeurIPS 2024 · similarity 0.88)
• Scrutinize What We Ignore: Reining In Task Representation Shift Of Context-Based Offline Meta Reinforcement Learning (ICLR 2025 · similarity 0.88)
• Adaptive Advantage-Guided Policy Regularization for Offline Reinforcement Learning (ICML 2024 · similarity 0.88)