Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 70 of the 71 references it cites were found in a registry or an index; 28 of the 32 sources its claims cite were checked for support.

XLand-MiniGrid: Scalable Meta-Reinforcement Learning Environments in JAX

Open access
Nikulin, Alexander, Kurenkov, Vladislav, Zisman, Ilya, Agarkov, Artem, Sinii, Viacheslav, Kolesnikov, Sergey · 2024 · Advances in Neural Information Processing Systems 37

Abstract

Inspired by the diversity and depth of XLand and the simplicity and minimalism of MiniGrid, we present XLand-MiniGrid, a suite of tools and grid-world environments for meta-reinforcement learning research. Written in JAX, XLand-MiniGrid is designed to be highly scalable and can potentially run on GPU or TPU accelerators, democratizing large-scale experimentation with limited resources. Along with the environments, XLand-MiniGrid provides pre-sampled benchmarks with millions of unique tasks of varying difficulty and easy-to-use baselines that allow users to quickly start training adaptive agents. In addition, we have conducted a preliminary analysis of scaling and generalization, showing that our baselines are capable of reaching millions of steps per second during training and validating that the proposed benchmarks are challenging. XLand-MiniGrid is open-source and available at https://github.com/dunnolab/xland-minigrid.

Overview

CoChat Summary

TL;DR

XLand-MiniGrid is a JAX-based suite of grid-world environments and pre-sampled benchmarks for meta-reinforcement learning, designed to scale to GPU/TPU accelerators and millions of environment steps per second. The paper reports preliminary scaling, baseline training, and generalization experiments showing the benchmarks are challenging, especially on the 20th percentile and unseen tasks.

Key findings

  • XLand-MiniGrid provides four benchmark types—trivial, small, medium, and high—with one million unique rulesets each, plus three-million variants for medium and high.
  • Simulation throughput scales almost log-log linearly with parallel environments, saturating around \(2^{13}\) on a single A100 but reaching tens of millions of steps per second across multiple devices.
  • Training throughput for the recurrent PPO/RL² baseline saturates near one million steps per second on a single device and can reach almost six million with multiple devices; a large-scale high-1m run trained for one trillion transitions in 72 hours at ∼3.85M steps per second.
  • The proposed benchmarks remain challenging: on image observations the 20th percentile is near zero, and a large generalization gap remains even on the benchmark with minimal diversity when goal types are held out.

Methods

  • The library implements extensible rules and goals in JAX, with stateless environments, jit-compatible reset/step methods, and support for vectorization via `jax.vmap` and multi-device scaling via `jax.pmap`.
  • Benchmarks are procedurally generated as task trees of production rules and goals, with distractor rules/objects, and are pre-sampled and hosted for download; the initial release includes 38 registered environment variations.
  • Baselines are near-single-file recurrent PPO implementations for single-task environments and an RL² extension for meta-learning, trained on A100 GPUs with evaluation on 4096 tasks and 25 trials per task, mainly reporting the 20th percentile.

Limitations

  • The paper states that compared with full-scale XLand, it does not currently support multi-agent simulations, procedural generation of complex worlds, rules with multiple output entities, or goal composition; it also lacks some MiniGrid tiles such as lava or moving obstacles and does not provide natural-language encodings of rules and goals.
  • The paper states that JAX is more restrictive than PyTorch, can be difficult to debug, and is poorly suited to heterogeneous computations or conditional branching common in environment implementations; new rules and goals are hard-coded and not easily added without modifying source code.
  • The paper states that grid layouts must be chosen in advance and cannot be changed under jit-compilation, and that efficient procedural grid generation is left for future work; the text also makes evident that the baseline experiments are preliminary and that image-observation training introduces large overheads, fitting only 1024 parallel environments on a single GPU and yielding about 20k steps per second during training.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.52202/079017-1390

Cites

70 of 70 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

16 Valid · 3 Not valid · 2 Not checked · 68 Own results and interpretations

26 more claims · 1 with their cited sources checked

Create a free account to see all 89 claims and to verify the 1 reference still open.

See all claims — free account
  • It may state two separate claims as one. 29
  • Probably not a substantial claim. 38
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Citations

Cites

  • Attention Is All You Need

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404 · cited 1×

    “Further improvement may require changing the RNN architecture to a more hardware-friendly one, e.g., the Transformer (Vaswani et al., 2017), which we left for future work.”
    Method
  • PyTorch: An Imperative Style, High-Performance Deep Learning Library

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al. (2019) · arXiv · doi:10.48550/arxiv.1912.01703 · Cited by 31,473 · cited 2×

    “In general, while JAX is much more high-level than CUDA or Triton (Tillet et al., 2019), it is still much more restrictive than PyTorch (Paszke et al., 2019), can be difficult to debug, and is poorly suited to the heterogeneous computations or conditional branching that are common when implementing…”
    Background
  • Reinforcement Learning: An Introduction

    Sutton, R. S., & Barto, A. G. (1998) · IEEE Transactions on Neural Networks · doi:10.1109/tnn.1998.712192 · Cited by 27,225 · cited 1×

    “Grid world environments have a long history in RL research Sutton & Barto (2018), as they offer a number of attractive properties.”
    Background
  • Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation

    Cho, K., van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014) · Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.3115/v1/d14-1179 · Cited by 25,545 · cited 1×

    “While our implementation is greatly simplified in comparison to Ada (Team et al., 2023), it encap-sulates the main principles and is easy to understand and modify, e.g., swapping simple GRU (Cho et al., 2014) for more modern and efficient state space models (Lu et al., 2023).”
    Method
  • Proximal Policy Optimization Algorithms

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017) · arXiv · doi:10.48550/arxiv.1707.06347 · Cited by 10,625 · cited 6×

    “With the release of XLand-MiniGrid, we are providing near-single-file implementations of recurrent PPO (Schulman et al., 2017) for single-task environments and its extension to RL 2 (Duan et al., 2016; Wang et al., 2016) for meta-learning as baselines.”
    Influential
  • Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks

    Finn, C., Abbeel, P., & Levine, S. (2017) · arXiv · doi:10.48550/arxiv.1703.03400 · Cited by 8,512 · cited 1×

    “…meta-RL benchmarks have focused on tasks with simple distributions, such as bandits and 2D navigation (Wang et al., 2016; Duan et al., 2016; Finn et al., 2017), or few-shot control policy adaptation (e.g., MuJoCo (Zintgraf et al., 2019) or MetaWorld (Yu et al., 2020)), where the latent…”
    Background
  • Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation

    Cho, K., van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014) · arXiv · doi:10.48550/arxiv.1406.1078 · Cited by 3,317

  • Deep Reinforcement Learning That Matters

    Henderson, P., Islam, R., Bachman, P., Pineau, J., Precup, D., & Meger, D. (2018) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v32i1.11694 · Cited by 1,685 · cited 1×

    “Reinforcement learning (RL) is known to be extremely sample inefficient and prone to overfitting, sometimes failing to generalize to even subtle variations in environmental dynamics or goals (Ra-jeswaran et al., 2017; Zhang et al., 2018; Henderson et al., 2018; Alver & Precup, 2020).”
    Background
  • Dota 2 with Large Scale Deep Reinforcement Learning

    OpenAI, :, Berner, C., Brockman, G., Chan, B., Cheung, V., Dębiak, P., Dennison, C., Farhi, D., Fischer, Q., et al. (2019) · arXiv · doi:10.48550/arxiv.1912.06680 · Cited by 1,065 · cited 1×

    “…lagged behind (Li et al., 2023), whereas on-policy methods, while generally less sample efficient, can scale to enormous batch sizes of millions (Berner et al., 2019; Petrenko et al., 2020) and complete training much faster (Stooke & Abbeel, 2018; Shacklett et al., 2021) in wall clock time.”
    Background
  • OpenAI Gym

    Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., & Zaremba, W. (2016) · arXiv · doi:10.48550/arxiv.1606.01540 · Cited by 582 · cited 4×

    “…the timestep is a bit different from the previous designs by Lange (2022); Bonnet et al. (2023), but it allows for some simplifications, such as an auto-reset wrapper implementation based on the current or previous step (in the style of Gym (Brockman et al., 2016) or EnvPool (Weng et al., 2022)).”
    BackgroundMethodInfluential

Show all 71 references

Research works printed but not linked: 2

  1. Kapturowski, S., Ostrovski, G., Quan, J., Munos, R., and Dabney, W. Recurrent experience replay in distributed reinforcement learning. In International Conference on Learning Representations , 2018. URL https://api.semanticscholar.org/CorpusID:59345798 .
  2. Kirsch, L., Harrison, J., Freeman, C., Sohl-Dickstein, J., and Schmidhuber, J. Towards generalpurpose in-context learning agents. In NeurIPS 2023 Workshop on Generalization in Planning , 2023. URL https://openreview.net/forum?id=eDZJTdUsfe .

Cited by

See all 3

Metadata

Authors
Nikulin, A., Kurenkov, V., Zisman, I., Agarkov, A., Sinii, V., & Kolesnikov, S.
Venue
Advances in Neural Information Processing Systems 37
Published
2024
Pages
43809-43835
ISSN
1049-5258 (ISSN-L)
OpenAlex
W4415797730
Cited by
3
References
71
Citations by year
2026: 1 · 2024: 2
FWCI
0.91
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
  • neural information processing systems (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Reinforcement Learning in Robotics
  • Physical Sciences › Computer Science › Artificial Intelligence › Artificial Intelligence in Games
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
meta-reinforcement learning, JAX, benchmark environments, meta-optimization

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: XLand-MiniGrid: Scalable Meta-Reinforcement Learning Environments in JAX. https://app.cochat.ai/papers/10.52202/079017-1390
The paper
Nikulin, A., Kurenkov, V., Zisman, I., Agarkov, A., Sinii, V., & Kolesnikov, S. (2024). XLand-MiniGrid: Scalable Meta-Reinforcement Learning Environments in JAX. Advances in Neural Information Processing Systems 37. https://doi.org/10.52202/079017-1390

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.