Address scarcity of high-quality robot manipulation data for training generalist policies

Develop scalable data collection, curation, or learning approaches that alleviate the scarcity of high-quality robot manipulation demonstrations needed to train generalist robotic manipulation policies effectively.

Background

Training generalist policies benefits from large and diverse datasets, yet collecting high-quality robot demonstrations is slow and expensive. The authors combine internal and public data but still note scarcity of high-quality data as an open issue.

Solutions that reduce dependence on costly data or improve data quality at scale would strengthen the reliability and generalization of LBMs.

References

Despite progress in training generalist policies, challenges such as catastrophic forgetting, data heterogeneity, scarcity of high-quality data, multimodal fusion, handling dexterity, and maintaining real-time inference speed remain open research problems.

— A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation  (2507.05331 - Team et al., 7 Jul 2025) in Section 2.1, Related Work—Robot Learning at Scale

To mitigate the gap between gameplay and robot execution, inspired by prior works, we leave action-level supervision from game data to future work and instead extract transferable information from trajectories, specifically contact points and sub-goal states that mark task progress.

— From Gameplay to Policy: Towards Scalable Robot Data Collection via Gamified Robot-Free Interaction  (2609.18650 - Li et al., 16 Sep 2026) in Section 3, subsection “Game2Policy: Affordance Transfer for Few-Shot Policy Learning”

Modern simulators used for reinforcement learning (RL) cannot fully replicate the required contact complexity, while collecting dexterous demonstrations for imitation learning (IL) remains an open problem.

— Rapid Learning of Dexterous In-Hand Pen Writing through Real-Time Jacobian Estimation  (2609.11775 - Stewart et al., 10 Sep 2026) in Abstract; Section 1, Introduction

The expert library is predefined by skill semantics. Learning its structure and reducing the need for human demonstrations remain open directions.

— RoboCoach: World Models as Active Coaches for Compositional Robot Skills  (2609.39685 - Liu et al., 30 Sep 2026) in Conclusion, paragraph “Limitations”

We do not evaluate a 20-human-only condition, so this comparison does not isolate the incremental benefit of adding generated trajectories to those 20 human demonstrations.

— Bridging Frontier Reasoning and Robot Execution: From Autonomous Demonstration Generation to Dense Language Supervision  (2610.03615 - Kim et al., 2 Oct 2026) in Appendix, Section 'Human-Only and Mixed Demonstration Corpora' (also discussed in Section 'Limitations and Future Work')