Quality over Quantity: Demonstration Curation via Influence Functions for Data-Centric Robot Learning

Abstract: 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.

• The paper introduces QoQ, an influence-function method that ranks demonstrations by how strongly their normalized policy gradients align with successful validation behaviors.
• QoQ combines maximum influence scoring with trajectory-level selection and scalable gradient compression, achieving 99.2% simulation and 86.7% real-world success, up to 30 percentage points above baselines.
• The method remains effective on heterogeneous demonstrations and policy rollouts, but future work must address sub-trajectory selection, cross-embodiment data, computational cost, and offline reinforcement learning.

Motivation and problem statement

Teleoperated demonstration datasets contain human errors, operational constraints, and teleoperator variability, all of which degrade behavior cloning (BC) performance. Existing curation approaches rely on proxy quality metrics — similarity to expert demonstrations in latent space (Du et al., 2023), optical-flow similarity [FlowRetrieval], or mutual information between states and actions (Hejna et al., 12 Feb 2025) — none of which directly measure whether a training sample improves the learned policy. The paper proposes QoQ, which grounds data quality in a direct quantity: a sample's contribution to reducing loss on a small validation set of desirable demonstrations, estimated via influence functions [Koh & Liang 2017].

Method

QoQ operates on a policy $\pi_{\theta_{\tt tr}}$ trained via BC on the full dataset. For each state-action pair $(s,a)$, it computes the normalized gradient $g(s,a)$ of the log-likelihood (or flow-matching loss for GR00T N1) and defines:

$$\text{QoQ-score}(s,a) := \max_{(s',a') \in \mathcal{D}_{\tt val}} g(s',a')^\top g(s,a).$$

Two design choices distinguish QoQ from prior influence-based curation such as CUPID (Agia et al., 23 Jun 2025) and DataMIL (Dass et al., 14 May 2025):

• Maximum influence scoring: rather than summing or averaging gradient products over all validation transitions, QoQ takes the maximum over validation pairs. Because different validation transitions encode different behaviors (e.g., pick-and-place vs. screwing), averaging introduces noise; focusing on the most relevant validation pair yields more reliable scores.
• Trajectory-wise curation: per-pair scores are averaged within each trajectory, and whole trajectories are ranked. Naive state-action-wise selection disproportionately retains redundant behaviors (e.g., grasping moments), harming state coverage; trajectory-level selection preserves complete behavior sequences.

For scalability to billion-parameter VLAs, gradients are computed only on selected layers (excluding vision encoders), and gradient vectors are compressed with one-permutation one-random-projection (OPORP) while preserving dot-product structure. Influence estimation itself uses the Hessian-free TracIn-style first-order approximation with normalized gradients [Pruthi et al. 2020].

Experimental results

Curation accuracy. On Robomimic "Can" (100 successful / 100 failed trajectories, 10 held-out successes as validation) and real-robot banana grasping (60/40 split), QoQ substantially outperforms Behavior Retrieval and Flow Retrieval:

Method	Simulation	Real (banana grasping)
QoQ	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

Downstream success rates. Policies fine-tuned on QoQ-curated data reach 99.2% success in simulation versus 76.0% for the best baseline (Flow Retrieval), and 86.7% on real banana grasping versus 56.7% for Behavior Retrieval — improvements of up to 23.2% and 30.0%, respectively. In multi-object curation, where the training set contains only other objects (peach, mango, snack, gum) and validation is successful banana pick-and-place, QoQ-curated policies achieve 93.3% success while Behavior Retrieval collapses to 20%, suggesting VAE latent representations are distracted by object diversity. QoQ also shows the most consistent trajectory rankings across seeds (Kendall's W = 0.7713 vs. 0.3287 for Behavior Retrieval).

In-the-wild robustness. On 200 DROID pen/pencil trajectories spanning heterogeneous environments, viewpoints, and object locations, QoQ maintains the highest curation accuracy, whereas retrieval baselines suffer from training VAE encoders on heterogeneous visual inputs.

Policy rollouts as validation. Without any held-out teleoperation data, QoQ can use rollouts from an uncurated policy as validation: weighted differences between QoQ scores against successful and failed rollouts identify harmful trajectories. On a real cabinet-opening task (5 successful / 15 failed rollout trajectories collected), this procedure improves deployment success over training on all data.

Ablations. Both maximum influence scoring and trajectory-wise curation individually improve curation accuracy and downstream success rate. Notably, computing gradients from only the action head (83.6% accuracy) or backbone (82.7%) matches full-parameter computation (82.1%), supporting scalability claims. Success rates remain above baselines across curation budgets from 10 to 60 trajectories, though small budgets reduce coverage (36.7% at budget 10 vs. 86.7% at budget 60), so the paper recommends sweeping budget sizes in practice.

Limitations and open questions

The authors acknowledge that trajectory-level curation cannot retain only high-quality segments within a mixed-quality trajectory, motivating sub-trajectory granularity. Influence estimates remain approximate and computationally nontrivial despite layer restriction and OPORP compression. The method assumes shared embodiment between training and validation data; extension to cross-embodiment settings such as Open X-Embodiment remains untested. Finally, evaluation is confined to BC-style objectives, leaving application to offline RL objectives unexplored.

Conclusion

QoQ replaces heuristic proxy metrics for robot demonstration quality with a performance-grounded criterion based on influence functions, adapted to robot learning through maximum influence scoring and trajectory-wise aggregation. Across simulation, real-robot manipulation, and DROID-scale heterogeneous data, it consistently outperforms feature-similarity baselines in both curation accuracy and downstream policy success, while remaining tractable on modern VLAs through selective-layer gradient computation and random projection.