Learning from demonstrations has emerged as a promising paradigm for end-to-end robot control, particularly when scaled to diverse and large datasets. However, the quality of demonstration data, often collected through human teleoperation, remains a critical bottleneck for effective data-driven robot learning. Human errors, operational constraints, and teleoperator variability introduce noise and suboptimal behaviors, making data curation essential yet largely manual and heuristic-driven. In this work, we propose Quality over Quantity (QoQ), a grounded and systematic approach to identifying high-quality data by defining data quality as the contribution of each training sample to reducing loss on validation demonstrations. To efficiently estimate this contribution, we leverage influence functions, which quantify the impact of individual training samples on model performance. We further introduce two key techniques to adapt influence functions for robot demonstrations: (i) using maximum influence across validation samples to capture the most relevant state-action pairs, and (ii) aggregating influence scores of state-action pairs within the same trajectory to reduce noise and improve data coverage. Experiments in both simulated and real-world settings show that QoQ consistently improves policy performances over prior data selection methods.
Figures & tables
Fig. 1: Illustration of high-quality data. Our robot data curation method, QoQ, selects trajectories based on their direct contribution to policy performance , using influence functions [ 13 ] to quantify this impact. Specifically, we measure the similarity between gradients of validation data (blue) and those of training data; higher similarity (green) indicates that including a particular state-action pair will effectively reduce validation loss. By prioritizing these high-impact data points, we systematically identify and preserve the most valuable training data that drive performance improvement.
Fig. 2: Success rate for simulation and real robot experiments. QoQ outperforms all baselines in simulation and real robot experiments by detecting helpful trajectories. We report the mean and standard deviation across 5 runs (simulation) and 3 runs (real robot).
Fig. 3: Visualization of experiment environments across simulation and real robot setups, including a single task of grasping a banana and multi-object pick-and-place, and open cabinet task.
Method
Simulation
Real (Banana grasping)
QoQ (Ours)
99.4 ± 0.3
83.6 ± 0.8
Behavior Retrieval
67.8 ± 0.7
67.3 ± 0.7
Flow Retrieval
56.9 ± 0.2
57.7 ± 0.1
All Data
55.4
58.7
TABLE I: Curation accuracy (%). Percentage of state-action pairs from successful trajectories among all pairs in the curated dataset.
Fig. 4: Droid dataset curation accuracy (%). Compared to baselines, QoQ maintains high curation accuracy in DROID dataset, which consists of different domains and object locations.
Fig. 5: Open cabinet policy success rate.
Fig. 6: Ablation study of QoQ components in banana grasping experiment. Removing maximum influence scoring and trajectory-wise curation reduces curation accuracy and policy success rate. Error bars show mean ± stderr over 3 runs.
Layer Part
Backbone
Action Head
All
Curation Accuracy
82.7 ± 0.6
83.6 ± 1.3
82.1 ± 1.2
TABLE II: Curation accuracy across different influence computation layer parts in banana grasping experiment. Backbone denotes the VLM and vision encoder part used in GROOT N1 model.
Curation Budget
10
20
40
60
All data
Best baseline
Success Rate (%)
36.7 ± 17.6
63.3 ± 17.6
60.0 ± 10.0
86.7 ± 8.9
36.7 ± 27.3
56.7 ± 17.6
TABLE III: Success rates across curation budgets. QoQ consistently achieves higher success rates across a wide range of budgets compared to baselines. Best baseline refers to the Behavior Retrieval method.
Metric
QoQ
Behavior Retrieval
Flow Retrieval
Kendall’s W
0.7713
0.3287
0.7026
TABLE IV: Kendall’s W values across different methods. Higher indicates more consistent trajectory rankings.