polixirai/NeoRL2

Dataset Card for NeoRL‑2: Near Real‑World Benchmarks for Offline Reinforcement Learning

Dataset Summary

NeoRL-2 is a collection of seven near–real-world offline-RL datasets plus their evaluation simulators. This repo we provide the offline-RL dataset, while the simulators are in https://github.com/polixir/NeoRL2.

Each task injects one or more realistic challenges—delays, exogenous disturbances, global safety constraints, traditional rule-based data, and/or severe data scarcity—into a lightweight control environment.

---

Dataset Details

Challenge	Brief description	Appears in
Delay	Long & variable observation-to-effect latency	Pipeline, Simglucose
External factors	State variables the agent cannot influence (e.g. wind, ground-friction)	RocketRecovery, RandomFrictionHopper, Simglucose
Global safety constraints	Hard limits that must never be violated	SafetyHalfCheetah
Rule-based behaviour policy	Trajectories from a PID or other deterministic controller	DMSD
Severely limited data	Tiny datasets reflecting expensive experimentation	Fusion, RocketRecovery, SafetyHalfCheetah

• Curated by: Polixir Technologies
• Paper: Gao et al. “NeoRL-2: Near Real-World Benchmarks for Offline Reinforcement Learning with Extended Realistic Scenarios”, arXiv:2503.19267 (2025)
• Repository (the environments for the datasets are in here): https://github.com/polixir/NeoRL2
• Task: offline / batch reinforcement learning

Uses

Direct Use

• Benchmarking offline-RL algorithms under near-deployment conditions
• Studying robustness to delays, safety limits, exogenous disturbances and data scarcity
• Developing data-efficient model-based or model-free methods able to outperform conservative behaviour policies

Loading example

from datasets import load_dataset

dmsd = load_dataset("polixir/neorl2", "DMSD", split="train")
state, action, reward, next_state, done = dmsd[0].values()

Out-of-Scope Use

• Online RL with unlimited interaction
• Safety-critical decision-making without extensive validation on the real system

---

Dataset Structure

Each Parquet row contains

Key	Type	Description
observations	float32[]	Raw observation vector (dim varies per task)
actions	float32[]	Continuous action taken by the behaviour policy
rewards	float32	Scalar reward
next_observations	float32[]	Observation at the next timestep
terminals	bool	True if episode ended (termination or safety)

Typical dataset sizes are ≈100 k transitions; Fusion, RocketRecovery and SafetyHalfCheetah are smaller by design.

---

Baseline Benchmark

Normalised return (0 – 100)

Task	Data	BC	CQL	EDAC	MCQ	TD3BC	MOPO	COMBO	RAMBO	MOBILE
Pipeline	69.25	68.6 ± 13.4	81.1 ± 8.3	72.9 ± 4.6	49.7 ± 7.4	81.9 ± 7.5	−26.3 ± 92.7	55.5 ± 4.3	24.1 ± 74.4	65.5 ± 4.1
Simglucose	73.9	75.1 ± 0.7	11.0 ± 3.4	8.1 ± 0.3	29.6 ± 5.7	74.2 ± 0.4	34.6 ± 28.1	23.2 ± 2.5	10.8 ± 0.9	9.3 ± 0.2
RocketRecovery	75.3	72.8 ± 2.5	74.3 ± 1.4	65.7 ± 9.8	76.5 ± 0.8	79.7 ± 0.9	−27.7 ± 105.6	74.7 ± 0.7	−44.2 ± 263.0	43.7 ± 17.5
RandomFrictionHopper	28.7	28.0 ± 0.3	33.0 ± 1.2	34.7 ± 1.3	31.7 ± 1.3	29.5 ± 0.7	32.5 ± 5.8	34.1 ± 4.7	29.6 ± 7.2	35.1 ± 0.5
DMSD	56.6	65.1 ± 1.6	70.2 ± 1.1	78.7 ± 2.3	77.8 ± 1.2	60.0 ± 0.8	68.2 ± 0.7	68.3 ± 0.4	76.2 ± 1.9	64.4 ± 0.8
Fusion	48.8	55.2 ± 0.3	55.9 ± 1.9	58.0 ± 0.7	49.7 ± 1.1	54.6 ± 0.8	−11.6 ± 22.2	55.5 ± 0.3	59.6 ± 5.0	5.0 ± 7.1
SafetyHalfCheetah	73.6	70.2 ± 0.4	71.2 ± 0.6	53.1 ± 11.1	54.7 ± 4.3	68.6 ± 0.4	23.7 ± 24.3	57.8 ± 13.3	−422.4 ± 307.5	8.7 ± 3.9

How often do algorithms beat the behaviour policy?

Margin	BC	CQL	EDAC	MCQ	TD3BC	MOPO	COMBO	RAMBO	MOBILE
≥ 0	3	4	4	4	6	2	3	3	2
≥ +3	2	4	4	2	4	2	3	2	2
≥ +5	2	3	3	1	2	1	3	2	2
≥ +10	0	2	1	1	1	1	1	2	0

Key conclusions

• No baseline “solves” any task (score ≥ 95). Best result is TD3BC’s 81.9 on Pipeline.
• TD3BC is the most reliable algorithm, surpassing the data in 6 / 7 tasks and still leading at stricter margins.
• Model-based methods (MOPO, RAMBO, and MOBILE) are brittle, with large variance and occasional catastrophic divergence.
• DMSD is easiest: many algorithms exceed the behaviour policy by 20 + points thanks to simple PID data.
• SafetyHalfCheetah is hardest: every method trails the data due to strict safety penalties and limited samples.
• In general, model-free approaches show smaller error bars than model-based ones, underlining the challenge of learning accurate dynamics under delay, disturbance and scarcity.

---

Citation

@misc{gao2025neorl2,
  title   = {NeoRL-2: Near Real-World Benchmarks for Offline Reinforcement Learning with Extended Realistic Scenarios},
  author  = {Songyi Gao and Zuolin Tu and Rong-Jun Qin and Yi-Hao Sun and Xiong-Hui Chen and Yang Yu},
  year    = {2025},
  eprint  = {2503.19267},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG}
}

---

Contact

Questions or bug reports? Please open an issue on the NeoRL-2 GitHub repo.