NeoRL-2

Source

• Dataset metadata snapshot: neorl2-2025
• Metadata JSON: metadata.json
• Official GitHub: https://github.com/polixir/NeoRL2
• Official Hugging Face dataset: https://huggingface.co/datasets/polixirai/NeoRL2
• arXiv preprint: https://arxiv.org/abs/2503.19267

Core Claim

NeoRL-2 is a near-real-world offline reinforcement-learning benchmark with explicit transition tuples and evaluation simulators. It targets practical offline RL difficulties that are underrepresented in simpler benchmarks: delays, exogenous factors, safety constraints, rule-based behavior policies, conservative data, and limited data.

Dataset Notes

• The paper and GitHub README describe seven tasks: Pipeline, Simglucose, RocketRecovery, RandomFrictionHopper, DMSD, Fusion, and SafetyHalfCheetah.
• The GitHub interface returns obs, next_obs, action, reward, done, and index.
• The Hugging Face parquet artifact uses observations, actions, rewards, next_observations, and terminals.
• Datasets are generated by online RL algorithms or PID policies, then suboptimal policies with returns from 50% to 80% of expert return are selected.
• Hugging Face reports 980848 rows and about 130 MB total file size.

Task Shapes

Task	Observation shape	Action shape	Done flag	Max timesteps
Pipeline	52	1	false	1000
Simglucose	31	1	true	480
RocketRecovery	7	2	true	500
RandomFrictionHopper	13	3	true	1000
DMSD	6	2	false	100
Fusion	15	6	false	100
SafetyHalfCheetah	18	6	false	1000

Action-Time-Series Notes

NeoRL-2 is a clean action-conditioned trajectory source:

observation_t + action_t -> reward_t + observation_{t+1} + terminal_t

This makes it better aligned with action-conditioned world-model training than logged decision datasets that only expose one-step outcomes. The harder part is that the benchmark intentionally includes delayed effects, external factors, conservative behavior policies, and safety constraints.

Gotchas

• The paper is an arXiv preprint from 2025; use it as a current benchmark artifact, not as peer-reviewed settled evidence.
• The tasks are simulated to reflect practical issues; they are not direct real-world business data.
• The paper reports that current baselines often fail to significantly improve over the behavior policy, and no reported baseline reaches the paper’s solved threshold.
• Hugging Face config metadata lists Salespromotion and Simglucose-high in addition to the seven paper/GitHub tasks.
• GitHub says datasets are CC BY 4.0 and code is Apache 2.0, while Hugging Face frontmatter marks the dataset repo as apache-2.0.

Foundation TSFM Relevance

Agenda slot	Verdict	Evidence	Missing pieces
Causal structure, counterfactuals, and control	partially closes	Provides explicit observation_t, action_t, reward, observation_{t+1}, and terminal signals across seven non-vision control tasks.	Simulators are benchmark approximations; no direct real-world deployment data.
Benchmarks: what level of modeling is tested?	partially closes	Stresses delayed effects, external factors, safety constraints, rule-based behavior policies, conservative data, and limited data.	Needs TSFM-native model comparisons and benchmark protocols beyond offline RL baselines.
Time representation and irregular event streams	adjacent	Pipeline and Simglucose explicitly test delay; trajectories have finite horizons and termination signals.	Mostly fixed simulator step interfaces rather than irregular event streams.
Context interface: channel context and general context	adjacent	Task identity and environment-specific properties define different control domains and constraints.	No unified typed context schema for cross-domain transfer.

Links Into The Wiki

• NeoRL-2
• Action-Conditioned Time-Series Datasets
• Time-Series Benchmark Hygiene
• World Models
• Foundation Time-Series Model Research Agenda