Robots That Learn From Old Data: Review Maps Offline Reinforcement Learning
Overview
The article explores the misunderstanding of data volume in the field of embodied intelligence, emphasizing the importance of data quality and efficiency. The author points out that simply accumulating data volume does not effectively improve model performance, but rather focuses on data diversity, information density, and the model's generalization and robustness capabilities. The article also introduces the practical approaches of Qiongche Intelligence in unbody data collection and world model construction, as well as how data loops and training loops enable large-scale deployment.
From Chinese robotics — Robot learning research
Event evidence and corrections
0 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
Reported quantity · required data hours: 10000000 hours · Basis not reported
“国金证券今年7月的报告估算，要训练一个具备实际作业能力的具身智能模型，行业至少需要1000万小时的多模态交互数据。”
Exact source · revision 2Source owner not reported
Reported quantity · data hours needed: 10000000 hours · Basis not reported
“国金证券今年7月的报告估算，要训练一个具备实际作业能力的具身智能模型，行业至少需要1000万小时的多模态交互数据。”
Exact source · revision 2Source owner not reported
Reported quantity · data hours needed for practical operation: 10000000 hours · Basis not reported
“国金证券今年7月的报告估算，要训练一个具备实际作业能力的具身智能模型，行业至少需要1000万小时的多模态交互数据。”
Exact source · revision 2Source owner not reported
Reported quantity · Reported success rate: 40 percent · Basis not reported · Differing source assertions
“tic constraints such as limited operator skill, scripted motion or safety monitors that cut episodes short. As reference points, the authors cite the D4RL pen-cloned variant, with roughly 5,000 transitions and a success rate below 40 percent, as a borderline case, and the relocate-cloned variant, with a success rate below 10 percent, as severely data-limited. Why does scarcity matter so much? Offline reinforcement le”
Exact source · revision 2Source owner not reported
Reported quantity · Reported success rate: 10 percent · Basis not reported · Differing source assertions
“tic constraints such as limited operator skill, scripted motion or safety monitors that cut episodes short. As reference points, the authors cite the D4RL pen-cloned variant, with roughly 5,000 transitions and a success rate below 40 percent, as a borderline case, and the relocate-cloned variant, with a success rate below 10 percent, as severely data-limited. Why does scarcity matter so much? Offline reinforcement le”
Exact source · revision 2Source owner not reported
Reported quantity · normalized returns: 86 other · Basis not reported · Differing source assertions
“mixed-quality data, CQL achieves normalized returns of approximately 86 to 95, whereas BCQ achieves 55 to 75 under comparable settings. IQL exceeds 85 on antmaze-medium-play, outperforming TD3+BC at roughly 48. On NeoRL’s low-data split, CQL degrades fewer than 10 normalized-return points from its full-data baseline while AWAC degrades more than 20 points. MOPO and MOReL surpass behavior-constrained baselines by 5 to”
Exact source · revision 2Source owner not reported
Reported quantity · normalized returns: 55 other · Basis not reported · Differing source assertions
“mixed-quality data, CQL achieves normalized returns of approximately 86 to 95, whereas BCQ achieves 55 to 75 under comparable settings. IQL exceeds 85 on antmaze-medium-play, outperforming TD3+BC at roughly 48. On NeoRL’s low-data split, CQL degrades fewer than 10 normalized-return points from its full-data baseline while AWAC degrades more than 20 points. MOPO and MOReL surpass behavior-constrained baselines by 5 to”
Exact source · revision 2Source owner not reported
Reported quantity · normalized returns: 85 other · Basis not reported · Differing source assertions
“mixed-quality data, CQL achieves normalized returns of approximately 86 to 95, whereas BCQ achieves 55 to 75 under comparable settings. IQL exceeds 85 on antmaze-medium-play, outperforming TD3+BC at roughly 48. On NeoRL’s low-data split, CQL degrades fewer than 10 normalized-return points from its full-data baseline while AWAC degrades more than 20 points. MOPO and MOReL surpass behavior-constrained baselines by 5 to”
Exact source · revision 2Source owner not reported
Reported quantity · normalized returns: 48 other · Basis not reported · Differing source assertions
“mixed-quality data, CQL achieves normalized returns of approximately 86 to 95, whereas BCQ achieves 55 to 75 under comparable settings. IQL exceeds 85 on antmaze-medium-play, outperforming TD3+BC at roughly 48. On NeoRL’s low-data split, CQL degrades fewer than 10 normalized-return points from its full-data baseline while AWAC degrades more than 20 points. MOPO and MOReL surpass behavior-constrained baselines by 5 to”
Exact source · revision 2Source owner not reported
Evidence dependency changes (7)
- 2026-10-30T19:53:09.512Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T19:52:41.536Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T19:45:53.251Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T19:45:38.633Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T19:30:54.490Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T19:30:40.656Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T17:29:18.392Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
Developments
- 2026-10-27 17:00 UTC · 1 reportsRobots That Learn From Old Data: Review Maps Offline Reinforcement LearningRobotics — Robot learning news discovery：Robots That Learn From Old Data: Review Maps Offline Reinforcement Learning Under Data Scarcity
- 2026-10-26 15:42 UTC · 1 reportsRobotics data scarcity and reproducibility challenges in embodied intelligenceChinese robotics — Dataset and collection discovery：Robotics Data Collection: Why Physical Interactions Are Hard to Replicate
- 2026-10-26 05:32 UTC · 1 reportsNoe-0: A VLA model trained entirely on non-corporate data for embodied AIChinese robotics — Robot learning research：The Most Dangerous Illusion in Embodied Intelligence, Targeting Those Who Believe in Millions of Hours
Report timeline
Follow attributed reports and material updates.
- Robotics — Robot learning news discoverySignalRobots That Learn From Old Data: Review Maps Offline Reinforcement Learning Under Data Scarcity
A new open-access review published in Discover Data examines how offline reinforcement learning performs in data-scarce robotic manipulation settings. The authors define data-limited scenarios based on five criteria, such as limited transition tuples or reward sparsity, and analyze six algorithmic families, including conservative value-based methods, behavior-constrained approaches, and model-based techniques. The study evaluates their effectiveness on benchmark datasets like D4RL and NeoRL, noting performance variations under different data conditions.
- Chinese robotics — Dataset and collection discoveryRobotics Data Collection: Why Physical Interactions Are Hard to Replicate
The article explores the challenges of training humanoid robots by highlighting the lack of available data and the difficulty of replicating physical interactions across different robot bodies and environments. It discusses the shift from data abundance to data scarcity, the rise of data factories, and the need for new data infrastructure and standards to support embodied AI development.
- Chinese robotics — Robot learning researchThe Most Dangerous Illusion in Embodied Intelligence, Targeting Those Who Believe in Millions of Hours
The article explores the misunderstanding of data volume in the field of embodied intelligence, emphasizing the importance of data quality and efficiency. The author points out that simply accumulating data volume does not effectively improve model performance, but rather focuses on data diversity, information density, and the model's generalization and robustness capabilities. The article also introduces the practical approaches of Qiongche Intelligence in unbody data collection and world model construction, as well as how data loops and training loops enable large-scale deployment.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).