Research map / Reinforcement learning / Visual and representation RL

Meta-World+: An Improved, Standardized, RL Benchmark

NeurIPS 2025 · Reinforcement learning › Visual and representation RL · approach: Task-inference meta-RL · 5 citations (OpenAlex)

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.

Re-engineers Meta-World into a standardized, Gymnasium-compatible release that preserves both V1 and V2 reward functions, and empirically shows that reward-version inconsistencies distort multi-task and meta-RL comparisons.

Problem

Meta-World has undocumented versioning changes (V1 vs V2 dense rewards) that make results across papers incomparable, and its dependencies on OpenAI gym and Mujoco-Py are unsupported.

Approach

Instead of introducing another reward version, the new release explicitly keeps both V1 and V2 reward functions selectable through a standardized API, adds MT25/ML25 task sets and custom task-set creation, and moves to the Gymnasium API with the Mujoco bindings; it also re-runs MTRL and meta-RL baselines across both reward versions to quantify the discrepancies.

Key results

• On MT10, PaCo scores 26.2 on V1 rewards versus 73.6 on V2, and MOORE scores 61.4 on V1 versus 83.2 on V2, showing reward-version sensitivity.
• On MT50, all tested multi-task methods score higher on V2 than V1; PaCo rises from 18.6 (V1) to 58.4 (V2).
• Own PCGrad replication on MT10 V1 gives 70.5 versus 90 reported in the original publication; SM, PaCo and MOORE closely match published numbers.

Measured effects reported in the paper

Surprising findings

• PCGrad could not be replicated to its published MT10 and MT50 numbers, likely due to a no-longer-available Meta-World version.
• MAML-V1, MAML-V2 and RL2-V2 show statistically no difference on ML10/ML45 despite the reward version changes.

Limitations

• Authors note the benchmark lacks language-instruction, multi-stage, and heterogeneous-reward task structure, which limits its scope.
• Meta-RL results on ML10/ML45 are modest, leaving a gap between ML1 parametric variation and non-parametric task distributions.
• Several baseline numbers (e.g., PCGrad) could not be reproduced because the original Meta-World version is no longer available.

Experimental setup

Datasets: Meta-World MT10 Meta-World MT50 Meta-World ML10 Meta-World ML45 Meta-World MT25 Meta-World ML25

Benchmarks: Meta-World MT10 Meta-World MT50 Meta-World ML10 Meta-World ML45

Metrics: mean success rate interquartile mean (IQM) 95% confidence interval

Baselines: MTMHSAC Soft Modularization PaCo MOORE PCGrad MAML RL2 AMAGO-2

Builds on: Meta-World PCGrad MTMHSAC MAML RL2

Closest papers by meaning

• AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.91)
• MAMBA: an Effective World Model Approach for Meta-Reinforcement Learning (ICLR 2024 · similarity 0.90)
• MetaCARD: Meta-Reinforcement Learning with Task Uncertainty Feedback via Decoupled Context-Aware Reward and Dynamics Components (AAAI 2024 · similarity 0.90)
• DynaMITE-RL: A Dynamic Model for Improved Temporal Meta-Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
• Meta-Reinforcement Learning with Universal Policy Adaptation: Provable Near-Optimality under All-task Optimum Comparator (NeurIPS 2024 · similarity 0.89)
• Meta-Reinforcement Learning via Exploratory Task Clustering (AAAI 2024 · similarity 0.88)