Summary
This paper releases a 1,500-hour bimanual household manipulation dataset (531.7 hours of real-robot teleoperation across 32,518 trajectories, plus roughly 1,000 hours of UMI-collected demonstrations) and uses it to train XR-2, a 5B-parameter VLA model, via a purpose-built high-throughput data pipeline and multi-stage training paradigm. It then studies two scaling axes explicitly: growing the amount of expert demonstration data, and post-training with DAgger-style on-policy correction data collected from real-time human interventions.
Key Contributions
- A large, task-diverse bimanual household manipulation dataset combining teleoperation and UMI-collected demonstrations at a scale (1,500 hours) uncommon outside a handful of industrial labs.
- A high-throughput data pipeline and multi-stage training paradigm engineered specifically for data efficiency at this scale, reported to retain favorable training efficiency despite the corpus size.
- A controlled empirical study of two distinct scaling axes — raw expert demonstration volume and on-policy DAgger correction volume — showing consistent, monotonic success-rate improvement across both, rather than just reporting a single scaled-up end result.
Strengths
- Explicitly isolating demonstration-volume scaling from correction-volume scaling gives the field a rare direct comparison of “more offline data” versus “targeted online correction data,” which is directly relevant to budgeting data-collection effort.
- Combining teleoperation with UMI (handheld gripper) data is a practical way to increase throughput and diversity beyond what teleoperation alone can achieve at this scale.
- The DAgger-correction results are a concrete, large-scale data point for the value of human-in-the-loop post-training, complementing existing DAgger-style methods in this space with far larger data volume than typical proof-of-concept studies.
Weaknesses
- Bimanual household manipulation is a specific task family; it’s unclear how the observed scaling trends generalize to other domains (industrial, outdoor, dexterous single-arm tasks).
- The paper doesn’t appear to compare against alternative uses of the same collection budget (e.g., RL fine-tuning or simulation-augmented data) to establish whether DAgger correction is the most cost-effective use of marginal data-collection effort versus other post-training strategies.
- A 5B-parameter model and purpose-built pipeline represent significant infrastructure investment that may not be reproducible outside well-resourced labs, limiting independent verification of the scaling claims.
Open Questions
- Does the observed scaling trend (steady, consistent improvement) show diminishing returns beyond 1,500 hours, or is the corpus still in the steep part of the scaling curve?
- How does the relative value of demonstration data versus DAgger correction data shift as the base policy becomes more competent?
- Would the same multi-stage training paradigm transfer to other embodiments, or is it tuned specifically for this bimanual platform?
Significance
One of the largest reported bimanual manipulation corpora with an accompanying controlled scaling study, this paper provides useful empirical grounding for a question the field has mostly answered anecdotally: whether to invest additional data-collection budget in more offline demonstrations or in targeted on-policy human corrections.