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 accessAbstract
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
Cites
Statistics
Verification
Claims the paper makes
16 Valid · 3 Not valid · 2 Not checked · 68 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“a single training of the Ada agent (Team et al., 2023) takes five weeks”
What the sources say
- Supportedcites Team et al., 2023 source
Appendix D.2 (p. 40) says training the single-agent version of AdA from scratch took approximately five weeks of wall-clock time. That total includes one week to train the teacher and four weeks to train AdA. (LLM, conf=0.90)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“most recent works on adaptive agents (Laskin et al., 2022; Lee et al., 2023; Lu et al., 2023; Norman & Clune, 2023) avoid complex environments in favor of more simplistic ones”
What the sources say
- Partly supportedcites Laskin et al., 2022 sourceExperimental — accuracy not yet measured
Section 4.1 (pp. 5–6) describes evaluations in a 10-arm bandit and small 2D grid environments, supporting the point that this work uses relatively simple settings. But it also evaluates a partially observable 3D visual DMLab Watermaze, so the paper does not support saying it avoids complex environments altogether. It cannot establish what most recent works do. (LLM, conf=0.84)
- Supportedcites Lee et al., 2023 source
The Experimental Setup (section 5.1) describes this paper’s test environments as a 10 × 10 room and a single-room navigation task with four colored boxes. The Miniworld section further notes that the latter has only four distinct tasks. This supports citing the paper as an example of adaptive-agent work using relatively simple environments. (LLM, conf=0.87)
- Partly supportedcites Wang et al., 2021 sourceExperimental — accuracy not yet measured
Related work (p. 3) says the most popular meta-RL benchmarks use relatively simple task parameterizations, supporting the claim's general contrast. It also describes environments designed for greater complexity, so it does not establish that most recent work on adaptive agents avoids complex environments. The 2021 paper cannot assess the cited 2022–2023 works. (LLM, conf=0.84)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“One way to address these shortcomings are meta-RL approaches, where adaptive agents are pre-trained on diverse task distributions to significantly increase sample efficiency on new problems”
What the sources say
- Supportedcites Wang et al., 2016 source
Section 2.2 (p. 2) describes training across interrelated tasks so the learned RL procedure can adapt rapidly to new task instances. The Conclusion (p. 14) says this procedure exploits shared task structure and can operate more efficiently than a general-purpose algorithm, supporting the sample-efficiency claim for related new problems. (LLM, conf=0.88)
- Supportedcites Duan et al., 2016 source
The Introduction (p. 2) says RL² optimizes an agent across a distribution of MDPs so it can adapt when deployed. The Abstract (p. 1) reports strong performance on previously unseen MDPs, and the tabular-MDP results (p. 6) report a large advantage over comparison methods when only a few episodes are available. The evidence concerns new tasks of the kinds used in training, not arbitrary unrelated problems. (LLM, conf=0.88)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Note that training Ada (Team et al., 2023) for such a duration would not be feasible. 7”
- Also stated in: § 5 Related Work · p. 9 · plain
“According to Appendix D.2 from Team et al. (2023), training Ada for 25 B transitions took over 1 week on 64 Google TPUv3.”
What the sources say
- Partly supportedcites Team et al., 2023 sourceExperimental — accuracy not yet measured
Appendix D.2 (p. 40) reports that training a single-agent version of AdA from scratch took approximately five weeks on 64 TPUv3 devices. That supports the claim that substantially longer training would be costly, but the paper does not assess one trillion transitions or say that training for that duration would be infeasible. (LLM, conf=0.78)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 5 Related Work · p. 9 · plain
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Unlike other RL subfields, meta-RL lacked environments that were both non-trivial and computationally efficient.”
What the sources say
- Could not verifycites Team et al., 2021 sourceExperimental — accuracy not yet measured
The Related Work passage mentions a curriculum method for meta-RL, but does not assess whether meta-RL lacked suitable environments. Searches for non-trivial, computationally efficient meta-RL environments found no passage in our copy making that comparison with other RL subfields. (LLM, conf=0.74)
- Could not verifycites Team et al., 2023 sourceExperimental — accuracy not yet measured
The Related Work discussion describes challenging procedural environments for RL and smaller-scale meta-RL settings, but does not say that meta-RL lacked environments combining non-trivial tasks with computational efficiency. Searches for “computationally efficient,” “non-trivial,” and simulation cost found no passage making that comparison. (LLM, conf=0.75)
- Partly supportedcites Wang et al., 2021 sourceExperimental — accuracy not yet measured
The Introduction (p. 2) says previous meta-RL environments tended to trade off accessibility and interestingness, and Related work (p. 3) contrasts simple benchmarks with more complex ones. But the paper defines accessibility as the ability to analyze a task distribution, not computational efficiency; it also does not establish the comparison with other RL subfields. (LLM, conf=0.87)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
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.
- It may state two separate claims as one. 29
- Probably not a substantial claim. 38
- The quote only partly states this claim. 2
Citations
Cites
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.”
MethodPyTorch: 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…”
BackgroundReinforcement 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.”
BackgroundLearning 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).”
MethodProximal 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.”
InfluentialModel-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…”
BackgroundLearning 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).”
BackgroundDota 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.”
BackgroundBrockman, 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)).”
BackgroundMethodInfluentialLearning to reinforcement learn
Wang, J. X., Kurth-Nelson, Z., Tirumala, D., Soyer, H., Leibo, J. Z., Munos, R., Blundell, C., Kumaran, D., & Botvinick, M. (2016) · arXiv · doi:10.48550/arxiv.1611.05763 · Cited by 579 · cited 3×
“One way to address these shortcomings are meta-RL approaches, where adaptive agents are pre-trained on diverse task distributions to significantly increase sample efficiency on new problems (Wang et al., 2016; Duan et al., 2016).”
IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
Espeholt, L., Soyer, H., Munos, R., Simonyan, K., Mnih, V., Ward, T., Doron, Y., Firoiu, V., Harley, T., Dunning, I., et al. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1802.01561 · Cited by 553 · cited 1×
“…many of the early breakthroughs on the Atari benchmark were driven by more efficient distributed experience collection (Horgan et al., 2018; Espeholt et al., 2018; Kapturowski et al., 2018), eventually reducing training time to just a few minutes (Stooke & Abbeel, 2018; Adamski et al.,…”
BackgroundIsaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning
Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., et al. (2021) · arXiv · doi:10.48550/arxiv.2108.10470 · Cited by 509 · cited 1×
“Porting the entire environment to the GPU, as was done in Isaac Gym (Makoviychuk et al., 2021), Megaverse (Petrenko et al., 2021) or Madrona (Shacklett et al., 2023), can remove this bottleneck, but has the disadvantage of being GPU-only.”
MethodRL$^2$: Fast Reinforcement Learning via Slow Reinforcement Learning
Duan, Y., Schulman, J., Chen, X., Bartlett, P. L., Sutskever, I., & Abbeel, P. (2016) · arXiv · doi:10.48550/arxiv.1611.02779 · Cited by 503 · cited 3×
“One way to address these shortcomings are meta-RL approaches, where adaptive agents are pre-trained on diverse task distributions to significantly increase sample efficiency on new problems (Wang et al., 2016; Duan et al., 2016).”
Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning
Yu, T., Quillen, D., He, Z., Julian, R., Narayan, A., Shively, H., Bellathur, A., Hausman, K., Finn, C., & Levine, S. (2019) · arXiv · doi:10.48550/arxiv.1910.10897 · Cited by 450 · cited 3×
“However, 800k on 4090 is still a good result, more than any other benchmark (like MetaWorld, Alchemy, etc) can achieve with such resources.”
Triton: an intermediate language and compiler for tiled neural network computations
Tillet, P., Kung, H. T., & Cox, D. (2019) · Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages · doi:10.1145/3315508.3329973 · Cited by 318 · 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…”
BackgroundD4RL: Datasets for Deep Data-Driven Reinforcement Learning
Fu, J., Kumar, A., Nachum, O., Tucker, G., & Levine, S. (2020) · arXiv · doi:10.48550/arxiv.2004.07219 · Cited by 310 · cited 1×
“Similar to Fu et al. (2020); Kurenkov et al. (2023), our benchmarks are hosted in the cloud and will be downloaded and cached the first time they are used.”
BackgroundBrax -- A Differentiable Physics Engine for Large Scale Rigid Body SimulationPreprint
Freeman, C. D., Frey, E., Raichuk, A., Girgin, S., Mordatch, I., & Bachem, O. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2106.13281 · Cited by 308 · cited 3×
“There have been many new JAX-based environments appearing recently (Freeman et al., 2021; Lange, 2022; Bonnet et al., 2023; Koyamada et al., 2023; Rutherford et al., 2023; Jiang et al., 2023; Frey et al., 2023; Lechner et al., 2023), each offering their own API variations.”
BackgroundMethodStructured State Space Models for In-Context Reinforcement Learning
Lu, C. X., Schroecker, Y., Gu, A., Parisotto, E., Foerster, J., Singh, S., & Behbahani, F. M. P. (2023) · Neural Information Processing Systems · doi:10.48550/arxiv.2303.03982 · Cited by 173 · cited 2×
“However, on multiple devices, scaling remains log-log linear without signs of saturation, easily reaching tens of millions of steps per second.”
BackgroundDiscovered Policy Optimisation
Lu, C., Kuba, J., Letcher, A., Metz, L., Witt, C. S. D., & Foerster, J. (2022) · neural information processing systems · doi:10.52202/068431-1197 · Cited by 157
Revisiting the Minimalist Approach to Offline Reinforcement Learning
Tarasov, D., Kurenkov, V., Nikulin, A., et al. (2023) · Neural Information Processing Systems · doi:10.48550/arxiv.2305.09836 · Cited by 151 · cited 1×
“The throughput can be improved even further by increasing the number of devices (see Figure 5e).”
BackgroundSelf-citationDiscovered Policy Optimisation
Lu, C., Kuba, J., Letcher, A., Metz, L., Witt, C. S. D., & Foerster, J. (2022) · Neural Information Processing Systems · doi:10.48550/arxiv.2210.05639 · Cited by 138 · cited 3×
“Overall, we used the default hyperparameters that are common in other PPO implementations, such as CleanRL (Huang et al., 2022) and PureJaxRL (Lu et al., 2022).”
Accelerated Methods for Deep Reinforcement LearningPreprint
Stooke, A., & Abbeel, P. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1803.02811 · Cited by 137 · cited 2×
“…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.”
BackgroundCleanRL: High-quality Single-file Implementations of Deep Reinforcement Learning Algorithms
Huang, S., Dossa, R. F. J., Ye, C., & Braga, J. (2021) · arXiv · doi:10.48550/arxiv.2111.08819 · Cited by 116
What Matters In On-Policy Reinforcement Learning? A Large-Scale Empirical Study
Andrychowicz, M., Raichuk, A., Stańczyk, P., Orsini, M., Girgin, S., Marinier, R., Hussenot, L., Geist, M., Pietquin, O., Michalski, M., et al. (2020) · arXiv · doi:10.48550/arxiv.2006.05990 · Cited by 115 · cited 1×
“Although it is not always optimal to maximize the number of steps per second, as this may reduce the sample efficiency (Andrychowicz et al., 2020), for demonstration purposes we searched over other parameters to maximize per device utilization.”
VariBAD: A Very Good Method for Bayes-Adaptive Deep RL via Meta-Learning
Zintgraf, L., Shiarlis, K., Igl, M., Schulze, S., Gal, Y., Hofmann, K., & Whiteson, S. (2019) · arXiv · doi:10.48550/arxiv.1910.08348 · Cited by 112 · cited 3×
“…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 component is reduced to a few parameters that control goal location or…”
BackgroundEnvPool: A Highly Parallel Reinforcement Learning Environment Execution Engine
Weng, J., Lin, M., Huang, S., Liu, B., Makoviichuk, D., Makoviychuk, V., Liu, Z.-C., Song, Y., Luo, T.-D., Jiang, Y.-K., et al. (2022) · Neural Information Processing Systems · doi:10.48550/arxiv.2206.10558 · Cited by 97 · cited 3×
“The most common approach would be to write the environment logic in low level languages (to bypass Python GIL) for asynchronous collection, as EnvPool Weng et al. (2022) does.”
BackgroundMethodTowards Generalization and Simplicity in Continuous Control
Rajeswaran, A., Lowrey, K., Todorov, E., & Kakade, S. (2017) · arXiv · doi:10.48550/arxiv.1703.02660 · Cited by 94 · 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).”
BackgroundA Dissection of Overfitting and Generalization in Continuous Reinforcement Learning
Zhang, A., Ballas, N., & Pineau, J. (2018) · arXiv · doi:10.48550/arxiv.1806.07937 · Cited by 88 · 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).”
BackgroundJumanji: a Diverse Suite of Scalable Reinforcement Learning Environments in JAXPreprint
Bonnet, C., Luo, D., Byrne, D., Surana, S., Abramowitz, S., Duckworth, P., Coyette, V., Midgley, L. I., Tegegn, E., Kalloniatis, T., et al. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2306.09884 · Cited by 82 · cited 6×
“However, most of them focus on single-task environments for robotics (Freeman et al., 2021), board games (Koyamada et al., 2023) or combinatorial optimization (Bonnet et al., 2023).”
BackgroundMethodResultInfluentialLearning to Model the World with LanguagePreprint
Lin, J., Du, Y., Watkins, O., Hafner, D., Abbeel, P., Klein, D., & Dragan, A. D. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2308.01399 · Cited by 80 · cited 2×
“For a language focused learning environment, we refer the reader to the recent HomeGrid (Lin et al., 2023) environment.”
BackgroundSome Considerations on Learning to Explore via Meta-Reinforcement Learning
Stadie, B. C., Yang, G., Houthooft, R., Chen, X., Duan, Y., Wu, Y., Abbeel, P., & Sutskever, I. (2018) · arXiv · doi:10.48550/arxiv.1803.01118 · Cited by 74 · cited 1×
“Despite the great variety of benefits of the existing grid world benchmarks, to the best of our knowledge, only the no longer maintained KrazyWorld (Stadie et al., 2018) focuses on meta-learning.”
BackgroundOpen-Ended Learning Leads to Generally Capable Agents
Open Ended Learning Team, Stooke, A., Mahajan, A., Barros, C., Deck, C., Bauer, J., Sygnowski, J., Trebacz, M., Jaderberg, M., Mathieu, M., et al. (2021) · arXiv · doi:10.48550/arxiv.2107.12808 · Cited by 56 · cited 3×
“Compared to the full-scale XLand (Team et al., 2021, 2023), we do not currently support multi-agent simulations, procedural generation of complex worlds, rules with multiple output entities, or goal composition.”
BackgroundBeBold: Exploration Beyond the Boundary of Explored RegionsPreprint
Zhang, T., Xu, H., Wang, X., Wu, Y., Keutzer, K., Gonzalez, J. E., & Tian, Y. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2012.08621 · Cited by 49 · cited 2×
“However, they pose a significant challenge even to modern RL methods, making it possible to test exploration (Zhang et al., 2020), language understanding and generalization (Hanjie et al., 2021; Zholus et al., 2022; Lin et al., 2023; Chevalier-Boisvert et al., 2023), as well as memory (Paischer et…”
BackgroundInfluentialJaxMARL: Multi-Agent RL Environments in JAXPreprint
Rutherford, A., Ellis, B., Gallici, M., et al. (2023) · arXiv.org · doi:10.48550/arxiv.2311.10090 · Cited by 45 · cited 1×
“There have been many new JAX-based environments appearing recently (Freeman et al., 2021; Lange, 2022; Bonnet et al., 2023; Koyamada et al., 2023; Rutherford et al., 2023; Jiang et al., 2023; Frey et al., 2023; Lechner et al., 2023), each offering their own API variations.”
The MineRL 2020 Competition on Sample Efficient Reinforcement Learning using Human PriorsPreprint
Guss, W. H., Castro, M. Y., Devlin, S., Houghton, B., Kuno, N. S., Loomis, C., Milani, S., Mohanty, S. P., Nakata, K., Salakhutdinov, R., et al. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2101.11071 · Cited by 40 · cited 1×
“This hierarchical structure is very similar to the way tasks are organized in the famous Minecraft (Guss et al., 2021) or the simpler Crafter (Hafner, 2021) benchmarks.”
MethodDistributed Deep Reinforcement Learning: Learn How to Play Atari Games in 21 minutes
Adamski, I., Adamski, R., Grel, T., Jędrych, A., Kaczmarek, K., & Michalewski, H. (2018) · Lecture notes in computer science · doi:10.1007/978-3-319-92040-5_19 · Cited by 40 · cited 1×
“…of the early breakthroughs on the Atari benchmark were driven by more efficient distributed experience collection (Horgan et al., 2018; Espeholt et al., 2018; Kapturowski et al., 2018), eventually reducing training time to just a few minutes (Stooke & Abbeel, 2018; Adamski et al., 2018) per game.”
BackgroundPetrenko, A., Huang, Z., Kumar, T., Sukhatme, G. S., & Koltun, V. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2006.11751 · Cited by 35 · 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.”
BackgroundHistory Compression via Language Models in Reinforcement LearningPreprint
Paischer, F., Adler, T., Patil, V., Bitto-Nemling, A., Holzleitner, M., Lehner, S., Eghbal-zadeh, H., & Hochreiter, S. (2022) · arXiv (Cornell University) · doi:10.48550/arxiv.2205.12258 · Cited by 29 · cited 1×
“…pose a significant challenge even to modern RL methods, making it possible to test exploration (Zhang et al., 2020), language understanding and generalization (Hanjie et al., 2021; Zholus et al., 2022; Lin et al., 2023; Chevalier-Boisvert et al., 2023), as well as memory (Paischer et al., 2022).”
BackgroundLarge Batch Simulation for Deep Reinforcement LearningPreprint
Shacklett, B., Wijmans, E., Petrenko, A., Savva, M., Batra, D., Koltun, V., & Fatahalian, K. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2103.07013 · Cited by 29 · 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.”
BackgroundIn-Context Reinforcement Learning for Variable Action SpacesPreprint
Sinii, V., Nikulin, A., Kurenkov, V., Zisman, I., & Колесников, С. В. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2312.13327 · Cited by 28 · cited 2×
“The recent wave of in-context reinforcement learning research (Laskin et al., 2022; Lee et al., 2023; Norman & Clune, 2023; Kirsch et al., 2023; Sinii et al., 2023; Zisman et al., 2023) also uses simple environments for evaluation, such as bandits, navigational DarkRoom & KeyToDoor, or MuJoCo with…”
BackgroundSelf-citationQ-Ensemble for Offline RL: Don't Scale the Ensemble, Scale the Batch Size
Nikulin, A., Kurenkov, V., Tarasov, D., et al. (2022) · arXiv.org · doi:10.48550/arxiv.2211.11092 · Cited by 23 · cited 1×
“Increasing the mini-batch size can also be beneficial in offline RL (Nikulin et al., 2022; Tarasov et al., 2023).”
BackgroundSelf-citationNovGrid: A Flexible Grid World for Evaluating Agent Response to NoveltyPreprint
Balloch, J. C., Lin, Z., Hussain, M., Srinivas, A., Wright, R., Peng, X., Kim, J. M., & Riedl, M. O. (2022) · arXiv.org · doi:10.48550/arxiv.2203.12117 · Cited by 20 · cited 1×
“For example, we can easily model novelty as a change in rules, goals or objects between episodes, similar to NovGrid (Balloch et al., 2022).”
ResultHuman-Timescale Adaptation in an Open-Ended Task Space
Adaptive Agent Team, Bauer, J., Baumli, K., Baveja, S., Behbahani, F., Bhoopchand, A., Bradley-Schmieg, N., Chang, M., Clay, N., Collister, A., et al. (2023) · arXiv · doi:10.48550/arxiv.2301.07608 · Cited by 19 · cited 10×
“According to Team et al. (2023), a full-scale XLand environment can use more than five rules.”
InfluentialAn Extensible, Data-Oriented Architecture for High-Performance, Many-World Simulation
Shacklett, B., Rosenzweig, L. G., Xie, Z., Sarkar, B., Szot, A., Wijmans, E., Koltun, V., Batra, D., & Fatahalian, K. (2023) · ACM Transactions on Graphics · doi:10.1145/3592427 · Cited by 19 · cited 1×
“Porting the entire environment to the GPU, as was done in Isaac Gym (Makoviychuk et al., 2021), Megaverse (Petrenko et al., 2021) or Madrona (Shacklett et al., 2023), can remove this bottleneck, but has the disadvantage of being GPU-only.”
MethodEmergence of In-Context Reinforcement Learning from Noise DistillationPreprint
Zisman, I., Kurenkov, V., Nikulin, A., Sinii, V., & Колесников, С. В. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2312.12275 · Cited by 18
Self-citationGriddly: A platform for AI research in games
Bamford, C. (2021) · Software Impacts · doi:10.1016/j.simpa.2021.100066 · Cited by 17 · cited 1×
“Other libraries, such as the popular MiniGrid (Chevalier-Boisvert et al., 2023) and Griddly (Bamford et al., 2020), are not scalable and extensible enough to cover meta-learning needs.”
BackgroundBenchmarking the Spectrum of Agent Capabilities
Hafner, D. (2021) · arXiv · doi:10.48550/arxiv.2109.06780 · Cited by 16 · cited 1×
“This hierarchical structure is very similar to the way tasks are organized in the famous Minecraft (Guss et al., 2021) or the simpler Crafter (Hafner, 2021) benchmarks.”
MethodGrounding Language to Entities and Dynamics for Generalization in Reinforcement Learning
Hanjie, A. W., Zhong, V., & Narasimhan, K. (2021) · arXiv · doi:10.48550/arxiv.2101.07393 · Cited by 15 · cited 1×
“…pose a significant challenge even to modern RL methods, making it possible to test exploration (Zhang et al., 2020), language understanding and generalization (Hanjie et al., 2021; Zholus et al., 2022; Lin et al., 2023; Chevalier-Boisvert et al., 2023), as well as memory (Paischer et al., 2022).”
BackgroundGigastep - One Billion Steps per Second Multi-agent Reinforcement Learning
Lechner, M., Yin, L.-H., Seyde, T., et al. (2023) · Neural Information Processing Systems · doi:10.52202/075280-0009 · Cited by 15 · cited 1×
“There have been many new JAX-based environments appearing recently (Freeman et al., 2021; Lange, 2022; Bonnet et al., 2023; Koyamada et al., 2023; Rutherford et al., 2023; Jiang et al., 2023; Frey et al., 2023; Lechner et al., 2023), each offering their own API variations.”
BackgroundIn-context Reinforcement Learning with Algorithm Distillation
Laskin, M., Wang, L., Oh, J., Parisotto, E., Spencer, S., Steigerwald, R., Strouse, D., Hansen, S., Filos, A., Brooks, E., et al. (2022) · arXiv · doi:10.48550/arxiv.2210.14215 · Cited by 14 · cited 2×
“…and this also has been pointed out by Wang et al. (2021), that such demanding requirements are the reason why most recent works on adaptive agents (Laskin et al., 2022; Lee et al., 2023; Lu et al., 2023; Norman & Clune, 2023) avoid complex environments in favor of more simplistic ones (e.g.,…”
BackgroundIGLU Gridworld: Simple and Fast Environment for Embodied Dialog Agents
Zholus, A., Skrynnik, A., Mohanty, S., et al. (2022) · arXiv.org · doi:10.48550/arxiv.2206.00142 · Cited by 13 · cited 1×
“…pose a significant challenge even to modern RL methods, making it possible to test exploration (Zhang et al., 2020), language understanding and generalization (Hanjie et al., 2021; Zholus et al., 2022; Lin et al., 2023; Chevalier-Boisvert et al., 2023), as well as memory (Paischer et al., 2022).”
BackgroundFrey, S. Y., Li, K., Nagy, P., Sapora, S., Lu, C. J. J., Zohren, S., Foerster, J., & Calinescu, A. (2023) · ACM International Conference on AI in Finance (ICAIF) · doi:10.1145/3604237.3626880 · Cited by 12 · cited 5×
“…2022; Lee et al., 2023; Norman & Clune, 2023; Kirsch et al., 2023; Sinii et al., 2023; Zisman et al., 2023) also uses simple environments for evaluation, such as bandits, navigational DarkRoom & KeyToDoor, or MuJoCo with random projections of observations (Lu et al., 2023; Kirsch et al., 2023).”
BackgroundMethodInfluentialAlchemy: A benchmark and analysis toolkit for meta-reinforcement learning agents
Wang, J. X., King, M., Porcel, N., Kurth-Nelson, Z., Zhu, T., Deck, C., Choy, P., Cassin, M., Reynolds, M., Song, F., et al. (2021) · arXiv · doi:10.48550/arxiv.2102.02926 · Cited by 10 · cited 5×
“However, 800k on 4090 is still a good result, more than any other benchmark (like MetaWorld, Alchemy, etc) can achieve with such resources.”
Influentialminimax: Efficient Baselines for Autocurricula in JAXPreprint
Jiang, M., Dennis, M., Grefenstette, E., & Rocktäschel, T. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2311.12716 · Cited by 9 · cited 1×
“There have been many new JAX-based environments appearing recently (Freeman et al., 2021; Lange, 2022; Bonnet et al., 2023; Koyamada et al., 2023; Rutherford et al., 2023; Jiang et al., 2023; Frey et al., 2023; Lechner et al., 2023), each offering their own API variations.”
BackgroundPodracer architectures for scalable Reinforcement Learning
Hessel, M., Kroiss, M., Clark, A., Kemaev, I., Quan, J., Keck, T., Viola, F., & van Hasselt, H. (2021) · arXiv · doi:10.48550/arxiv.2104.06272 · Cited by 8 · cited 2×
“This makes it possible to use the Anakin architecture (Hessel et al., 2021) and easily scale to multiple devices using the jax.pmap transformation.”
MethodSupervised Pretraining Can Learn In-Context Reinforcement Learning
Lee, J. N., Xie, A., Pacchiano, A., Chandak, Y., Finn, C., Nachum, O., & Brunskill, E. (2023) · arXiv · doi:10.48550/arxiv.2306.14892 · Cited by 7 · cited 2×
“The recent wave of in-context reinforcement learning research (Laskin et al., 2022; Lee et al., 2023; Norman & Clune, 2023; Kirsch et al., 2023; Sinii et al., 2023; Zisman et al., 2023) also uses simple environments for evaluation, such as bandits, navigational DarkRoom & KeyToDoor, or MuJoCo with…”
BackgroundMegaverse: Simulating Embodied Agents at One Million Experiences per SecondPreprint
Petrenko, A., Wijmans, E., Shacklett, B., & Koltun, V. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2107.08170 · Cited by 7 · cited 2×
“…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.”
BackgroundMethodFirst-Explore, then Exploit: Meta-Learning Intelligent Exploration
Norman, B., & Clune, J. (2023) · arXiv.org · doi:10.48550/arxiv.2307.02276 · Cited by 7 · cited 2×
“The recent wave of in-context reinforcement learning research (Laskin et al., 2022; Lee et al., 2023; Norman & Clune, 2023; Kirsch et al., 2023; Sinii et al., 2023; Zisman et al., 2023) also uses simple environments for evaluation, such as bandits, navigational DarkRoom & KeyToDoor, or MuJoCo with…”
BackgroundA Brief Look at Generalization in Visual Meta-Reinforcement LearningPreprint
Alver, S., & Precup, D. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2006.07262 · Cited by 5 · 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).”
BackgroundKatakomba: Tools and Benchmarks for Data-Driven NetHack
Kurenkov, V., Nikulin, A., Tarasov, D., et al. (2023) · Neural Information Processing Systems · doi:10.48550/arxiv.2306.08772 · Cited by 5 · cited 2×
“Similar to Fu et al. (2020); Kurenkov et al. (2023), our benchmarks are hosted in the cloud and will be downloaded and cached the first time they are used.”
BackgroundSelf-citationChevalier-Boisvert, M., Dai, B., Towers, M., de Lazcano, R., Willems, L., Lahlou, S., Pal, S., Castro, P. S., & Terry, J. (2023) · arXiv · doi:10.48550/arxiv.2306.13831 · Cited by 4 · cited 1×
“To avoid unnecessary overhead and standardize comparisons between algorithms and experiments, we pre-generated several benchmarks with up to three million unique rulesets.”
MethodCraftax: A Lightning-Fast Benchmark for Open-Ended Reinforcement Learning
Matthews, M., Beukman, M., Ellis, B., Samvelyan, M., Jackson, M., Coward, S., & Foerster, J. (2024) · arXiv · doi:10.48550/arxiv.2402.16801 · Cited by 4 · cited 1×
“The most similar to our work is Craftax (Matthews et al., 2024).”
Pgx: Hardware-accelerated parallel game simulation for reinforcement learning
Koyamada, S., Okano, S., Nishimori, S., et al. (2023) · arXiv.org · doi:10.48550/arxiv.2303.17503 · Cited by 3 · cited 3×
“There have been many new JAX-based environments appearing recently (Freeman et al., 2021; Lange, 2022; Bonnet et al., 2023; Koyamada et al., 2023; Rutherford et al., 2023; Jiang et al., 2023; Frey et al., 2023; Lechner et al., 2023), each offering their own API variations.”
Parallel $Q$-Learning: Scaling Off-policy Reinforcement Learning under Massively Parallel SimulationPreprint
Li, Z., Chen, T., Hong, Z.-W., Ajay, A., & Agrawal, P. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2307.12983 · Cited by 2 · cited 1×
“The throughput can be improved even further by increasing the number of devices (see Figure 5e).”
BackgroundEmergence of Collective Open-Ended Exploration from Decentralized Meta-Reinforcement Learning
Bornemann, R., Hamon, G., Nisioti, E., & Moulin-Frier, C. (2023) · HAL · doi:10.48550/arxiv.2311.00651 · Cited by 2 · cited 1×
“A similar procedure was also used in a concurrent work by Bornemann et al. (2023) but for multi-agent meta-learning.”
MethodLarge Batch Optimization for Deep Learning: Training BERT in 76 minutes
You, Y., Li, J., Reddi, S., Hseu, J., Kumar, S., Bhojanapalli, S., Song, X., Demmel, J., Keutzer, K., & Hsieh, C.-J. (2019) · arXiv · doi:10.48550/arxiv.1904.00962 · cited 1×
“Large-batch RL. Large batches are known to be beneficial in deep learning (You et al., 2019) and deep reinforcement learning is no exception.”
BackgroundDistributed Prioritized Experience Replay
Horgan, D., Quan, J., Budden, D., Barth-Maron, G., Hessel, M., van Hasselt, H., & Silver, D. (2018) · arXiv · doi:10.48550/arxiv.1803.00933 · cited 1×
“For example, many of the early breakthroughs on the Atari benchmark were driven by more efficient distributed experience collection (Horgan et al., 2018; Espeholt et al., 2018; Kapturowski et al., 2018), eventually reducing training time to just a few minutes (Stooke & Abbeel, 2018; Adamski et al.,…”
BackgroundEnvPool: A Highly Parallel Reinforcement Learning Environment Execution Engine
Weng, J., Lin, M., Huang, S., et al. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-1628
Research works printed but not linked: 2
- 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 .
- 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
Craftax: A Lightning-Fast Benchmark for Open-Ended Reinforcement Learning
Matthews, M., Beukman, M., Ellis, B., Samvelyan, M., Jackson, M., Coward, S., & Foerster, J. (2024) · arXiv · doi:10.48550/arxiv.2402.16801 · Cited by 4
Mixture of Meta-Policies for Cross-Environment Meta-Reinforcement Learning
Liu, X., Zeng, Q., Tang, C., & Lv, J. (2026) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3770854.3780332 · Cited by 1
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)
- arXiv
- 2312.12044
- OpenAlex
- W4415797730
- Cited by
- 3
- References
- 71
- Citations by year
- 2026: 1 · 2024: 2
- FWCI
- 0.91
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- 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
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.