Closing the Lab-to-Store Gap: A Data-Efficient Post-Training and Experience-Driven Learning VLA Framework for Retail Humanoids

arXiv:2607.20345 · cs.RO, cs.AI · Submitted 2026-07-22 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Closing the Lab-to-Store Gap".

Dev: Here's the gist of the paper. Closing the gap between benchmark performance and reliable real-world operation remains a central challenge for Vision-Language-Action (VLA) humanoid robots, which must handle execution errors, distribution shifts, and environmental variability <ref:2607.20345#pg2> The paper presents DEED (Data-Efficient Post-Training and Experience-Driven Learning), a systems-level approach evaluated on a supermarket chip-restocking task using a Unitree G1-Edu humanoid robot and the GR00T N1.6 foundation model <ref:2607.20345#pg2> DEED comprises three key components: (1) a data-efficient post-training pipeline with control-frequency alignment, data curation, task-relevant visual highlighting, and reduced VLA dependence <ref:2607.20345#pg2>, (2) a real-world study of experiencedriven refinement, adapted from RECAP via a text-based advantage prefix and a vision-language value function <ref:2607.20345#pg2>, and (3) a latent-space analysis tool for studying in- and out-ofdistribution behavior <ref:2607.20345#pg2>.

Rosa: Data Preparation for Deployment The Data-Efficient (DE-) post-training recipe focuses on practical engineering decisions necessary when finetuning a VLA out of the box <ref:2607.20345#pg2> This recipe includes several critical rules for data curation and pipeline design <ref:2607.20345#pg2>.

Dev: Key elements include: 1.

Rosa: Frequency Hierarchy, which establishes that recording or acting faster than the camera means learning from duplicated observations that have not changed between frames <ref:2607.20345#pg2> 2.

Paper summary: Rosa: So we've been looking at "Closing the Lab-to-Store Gap: A Data-Efficient Post-Training and Experience-Driven Learning VLA Framework for Retail Humanoids," which is essentially a blueprint for making these robots reliable outside the lab.

Dev: It boils down to using DEED, which combines a data efficient post training pipeline with an experience driven refinement stage and a latent space analysis tool.

Rosa: The authors show that by doing this, you can get good performance on tasks like chip restocking without needing massive amounts of perfectly labeled real world data upfront.

Dev: They demonstrate that the gains come from smart integration rather than just a new architectural design for the VLA itself.

Rosa: The paper suggests that for deployment in retail, you need to focus on how you handle execution errors and distribution shifts by using these specific post training techniques.

Dev: It points out that while refinement helps initially, it can degrade if the system tries to generalize too much during those self generated rollouts.

Rosa: The title itself highlights this gap we're talking about, showing that the challenge is bridging lab performance and reliable real world operation for these humanoid robots.

Dev: Ultimately, this work gives us a practical framework for deploying foundation model based humanoids in environments where you need them to perform tasks reliably in real store settings.

Rosa: That's the essence of DEED, showing a way to be data efficient and experience driven while keeping an eye on how the system behaves as it moves into the real world.

Conclusion: Rosa: So we're wrapping up our look at DEED, which is this whole system for making these humanoid robots actually work reliably in a store, not just on paper and in a lab setting.

Dev: Yeah, the title itself says "Closing the Lab-to-Store Gap," and that means they are tackling that big problem of getting lab results into real-world operation without needing endless perfect data.

Taro: I'm interested in the authors too, because this whole setup is about how you handle errors and shifts when the robot is actually out there doing its job.

Rosa: Right, so they developed this DEED framework with a bunch of components—post training for data efficiency, experience driven refinement, and a tool to check if the robot is getting confused by what it’s seen before or not.

Dev: The core idea is that you don't need massive labeled datasets to get a decent policy working on these complex tasks like restocking shelves.

Taro: I wonder how long this works in reality? Because when you refine things iteratively, the system starts generating its own data, and that’s where things usually break down.

Rosa: That’s the trade-off they show; one iteration helps a lot with success rate, but a second iteration can actually lower the success rate because it gets stuck in its own mistakes.

Dev: It suggests that refinement has limits when the task starts demanding more complex generalization than what the training data provides.

Taro: And this latent space monitoring tool they use is crucial for seeing *why* it's failing; it tells you if the robot is just doing a small wiggle or if it’s completely lost in a new area of operation.

Rosa: So, essentially, DEED gives us a systematic way to prepare and refine these models so they are more robust when deployed outside of perfect controlled environments.

Dev: It's about moving away from just building big models and focusing on the entire pipeline that goes from training data to deployment safety.

Taro: We'll talk more later about what happens when that monitoring tool flags a real failure during a live task, because that's where the most interesting stuff usually happens.

Roger Sala Siso, Tiago Silverio, Jakob Sand, Tran Nguyen Le

Equal Contribution 1 HIVE Robots, Denmark Department of Engineering Technology Technical University of Denmark

cs.RO, cs.AI

Submitted: 2026-07-22

Updated: 2026-11-02

Comments: 8 pages. Accepted for publication in 2026 IEEE-RAS 25th International Conference on Humanoid Robots (Humanoids)

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 81/100

The gist: Closing the gap between benchmark performance and reliable real-world operation remains a central challenge for Vision-Language-Action (VLA) humanoid robots…

Key concepts

Data-Efficient Post-Training Pipeline
This stage focuses on practical engineering decisions for finetuning VLA models. It includes rules like frequency hierarchy to avoid learning from redundant observations and visual highlighting of task-relevant areas. The goal is to create a competent system using minimal data and resources.
Experience-Driven Refinement
This stage adapts prior experience learning methods using a text prefix and a vision-language value function. It iteratively collects autonomous rollouts, allowing human operators to intervene only when necessary. This refines the policy by estimating the expected return of visual observations under instruction.
Latent-Space Analysis Tool
This tool measures how far the robot's current state is from previously seen states in a compressed latent space. It uses a Gaussian Mixture Model to approximate data density and calculates Mahalanobis distance to flag when the robot leaves its training distribution, enabling online monitoring for safety.
Frequency Hierarchy
This rule dictates that recording or acting faster than the camera means learning from observations that haven't changed between frames. This prevents the system from wasting effort on duplicated visual data during training.

Terminology

Summary

Closing the gap between benchmark performance and reliable real-world operation remains a central challenge for Vision-Language-Action (VLA) humanoid robots, which must handle execution errors, distribution shifts, and environmental variability <ref:2607. p. 020345#pg2> The paper presents DEED (Data-Efficient Post-Training and Experience-Driven Learning), a systems-level approach evaluated on a supermarket chip-restocking task using a Unitree G1-Edu humanoid robot and the GR00T N1. p. 06 foundation model <ref:2607. p. 020345#pg2> DEED comprises three key components: (1) a data-efficient post-training pipeline with control-frequency alignment, data curation, task-relevant visual highlighting, and reduced VLA dependence <ref:2607. p. 020345#pg2>, (2) a real-world study of experiencedriven refinement, adapted from RECAP via a text-based advantage prefix and a vision-language value function <ref:2607. p. 020345#pg2>, and (3) a latent-space analysis tool for studying in- and out-ofdistribution behavior <ref:2607. p. 020345#pg2>.

Data Preparation for Deployment

The Data-Efficient (DE-) post-training recipe focuses on practical engineering decisions necessary when finetuning a VLA out of the box <ref:2607.20345#pg2> This recipe includes several critical rules for data curation and pipeline design <ref:2607. p. 020345#pg2>. Key elements include:

  1. Frequency Hierarchy, which establishes that recording or acting faster than the camera means learning from duplicated observations that have not changed between frames <ref:2607. p. 120345#pg2>

  2. Data Curation Rules, such as ensuring balanced state coverage and retaining incidental recoveries to prevent noise and systemic bad patterns from compromising learning <ref:2607. p. 120345#pg2>.

  3. Highlighting Task-Relevant Visual Cues, where a large VLM highlights task-relevant image regions via segmentation masks propagated across frames, avoiding per-timestep inference <ref:2607. p. 220345#pg2>.

  4. Practical Tricks to Reduce VLA Dependence, suggesting a binary hand controller and action smoothing via Butterworth filtering to enhance applicability on humanoids <ref:2607. p. 220345#pg2>.

How the Training Signal is Built

The Experience-Driven (-ED) stage adapts the RECAP recipe to the GR00T architecture using a text-based instruction prefix and a vision-language only value function <ref:2607.20345#pg2> The refinement procedure follows an iterative loop where autonomous rollouts are collected, and human operators intervene only when necessary to correct erratic or failing behavior <ref:2607. p. 220345#pg2>.

The value function Vϕ(ot, l) is trained on the combined dataset of autonomous rollouts, teleoperation demonstrations, and human corrective interventions to estimate the expected return of each visual observation under instruction <ref:2607. p. 220345#pg2>. Advantage labels Aˆt are computed using a multi-step estimator from RECAP, balancing realized task progress over the horizon against the critic’s estimate of future improvement <ref:2607. p. 320345#pg2>. These continuous values are converted into binary labels by comparing them to a threshold τ chosen so that a target fraction ϵ = 0. p. 34 of the dataset is labeled positive <ref:2607. p. 420345#pg2>.

Where the Gains Come From

The results show that bridging the labto-store gap is primarily a systems integration challenge rather than an architectural one <ref:2607.20345#pg2> The Data-Efficient stage alone transforms a policy that fails under naive finetuning into a competent real-world system using only a single GPU <ref:2607. p. 020345#pg2>. A single Experience-Driven iteration improves both success rate and efficiency, raising success to 42% (21/50) while reducing per-bag time to 22. p. 537 s <ref:2607. p. 020345#pg2>. However, a second iteration shows a drop in success to 22% (11/50), which is hypothesized to stem from distributional effects during iterative refinement as self-generated rollouts come to dominate the training set <ref:2607. p. 620345#pg2>.

Monitoring for Robustness

The latent-space in/out-of-distribution analysis tool serves as a shared instrument across both DEED stages, measuring how far the robot’s current configuration lies from the states seen during data collection <ref:2607.20345#pg2>. States are compared in the latent space of the fine-tuned GR00T state encoder, and a Gaussian Mixture Model (GMM) with K components is fitted to approximate the empirical density <ref:2607. p. 420345#pg2>. A new encoded state is scored by its Mahalanobis distance to the nearest mixture component, which accounts for each mode’s covariance structure to separate benign variation from genuine departures off the demonstrated manifold <ref:2607. p. 420345#pg2>. This tool serves as a diagnostic that can be used online to detect when a deployed policy is leaving its training distribution <ref:2607. p. 720345#pg2>.

Conclusion on Refinement Dynamics

The degradation observed in the second iteration suggests that the benefit of refinement diminishes as a task demands more generalization, and the greater the variation, the harder it is for self-generated rollouts to cover it <ref:2607.20345#pg2>. The structural finding that refined policies tend to stop after a single placement rather than returning to their start configuration traces to the value function assigning higher value to terminal completion states, which suppresses reset transitions during training <ref:2607. p. 720345#pg2>. The framework rests on the assumption that similar states recur across episodes under consistent geometry, grounding advantage estimates in repeated observations <ref:2607. p. 620345#pg2>. The in-distribution tool illustrates the value of treating monitoring as a first-class part of the framework to close the loop that DEED is built around <ref:2607. p. 720345#pg2>.

REFERENCES

[1] M. J. Kim, K. Pertsch, S. p. 7 Karamcheti, T. p. 7 Xiao, A. Balakrishna, S. p. 7 Nair, R. Rafailov, E. p. 7 Foster, G. Lam, P., Sanketi., Q., Vuong., T., Kollar,, Burchfiel,, R Tedrake,, D Sadigh,, C Finn,, and S Levine,“Openvla: An open-source vision-language-action model,” 2024. p. 7[1]

[2] J. Bjorck, F. Castaneda, N Cherniadev, X Da, R Ding, L Fan. p. 7,˜F Yu, D Fox., F Hu., S Huang., J Jang,, Z Jiang,, J Kautz,, K Kundalia,, L E Lao,, Z Li,, Z Lin,, K Lin,, G Liu,, E Llontop,, L Magne,, A Mandlekar,, A Narayan, Y Zhu et al,“Gr00t n1: An open foundation model for generalist humanoid robots,” 2025. p. 7[2]

[3] K Black, N Brown, D Driess, A Esmail, M Equi, C Finn,, N Fusai,, L Groom,, K Hausman,, B Ichter,, S Jakubczak,, T Jones., L Ke,, S Levine. p. 7, A Li-Bell,, M Mothukuri. p. 7, S Nair., K Pertsch. p. 7, L X Shi., J Tanner, Q Vuong, A Walling, H Wang, and U Zhilinsky,“π0: A vision-language-action flow model for general robot control,” arXiv preprint arXiv:2410. p. 724164, 2024. p. 7[3]

[4] S Belkhale, Y Cui,, and D Sadigh,“Data quality in imitation learning,” Advances in Neural Information Processing Systems, vol. 36, 2023.[4]

[5] Physical Intelligence, A Z Ren., C Sharma., L X Shi., L Smith., J T Springenberg,, K Stachowicz,, W Stoeckle,, A Swerdlow,, J Tanner, M Torne, Q Vuong, A Walling, H Wang, B Williams,, S I Yoo,, L Yu,, U Zhilinsky, Z Zhou et al,“π∗0.6: a vla that learns from experience,” 2025. p. 7[5]

[6] E Hannus., M Malin., T N Le,, and V Kyrki,“Ia-vla: Input augmentation for vision-language-action models in settings with semantically complex tasks,” arXiv preprint arXiv:2509. p. 724768, 2025. p. 7[6]

[7] J Yang, K Chen, Z Li., S Wu., Y Zhao., L Ren., W Luo., C Shang,, M Zhi,, L Gao,, M Sun,, and H Cheng,“Bootstrapping imitation learning for long-horizon manipulation via hierarchical data collection space,” arXiv preprint arXiv:2505. p. 717389, 2025. p. 7[7]

[8] Z Hu, R Wu., N Enock, J Li., R Kadakia,, Z Erickson,, and A Kumar,“Rac: Robot learning for long-horizon tasks by scaling recovery and correction,” arXiv preprint arXiv:2509. p. 707953, 2025. p. 7[8]

[9] Y Xing, X Luo., J Xie., L Gao., H Shen,, and J Song,“Shortcut learning in generalist robot policies: The role of dataset diversity and fragmentation,” in Proceedings of The 9th Conference on Robot Learning (CoRL), vol. p. 7 305. PMLR, 2025, pp. p. 7 3239–3266. p. 7[9]

[10] Y Zhang, Y Qi., and X Zheng,“Experiences from benchmarking vision-language-action models for robotic manipulation,” arXiv preprint arXiv:2511. p. 711298, 2025. p. 7[10]

[11] K Lee, K Lee,, H Lee,, and J Shin,“A simple unified framework for detecting out-of-distribution samples and adversarial attacks,” in Advances in Neural Information Processing Systems (NeurIPS), 2018, pp. 7167–7177. p. 7[11]

[12] K Wang, Y Zheng, Y Zheng,, J Zhao,, and W Ding,“Learning high-frequency continuous action chunks in latent space,” 2026.[12]

[13] J Yu, H Liu, Q Yu., J Ren., C Hao,, H Ding,, G Huang,, G Huang,, Y Song, P Cai. p. 7, C Lu., and W Zhang,“Forcevla: Enhancing vla models with a force-aware moe for contact-rich manipulation,” arXiv preprint arXiv:2505. p. 722159, 2025. p. 7[13]

[14] H Zhang, W.-H Huang, Q Tong,, G Solak,, P Liu., K Zhang., S Liu., J Peters,, Y She,, and A Ajoudani,“Compliantvla-adaptor: Vlmguided variable impedance action for safe contact-rich manipulation,” arXiv preprint arXiv:2601. p. 715541, 2026. p. 7[14]

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Data-Efficient Post-Training Pipeline: Implement a control frequency hierarchy where recording frequency is set equal to inference control frequency and both must be less than or equal to camera frequency, ensuring Recording and inference control frequencies must match, since the policy learns the temporal semantics of the action space at whatever rate it was recorded at.

  2. Data Curation Rules: Ensure balanced state coverage by recording only efficient, successful behavior that avoids hesitations, corrections, backtracking, or inefficient paths, and retain incidental recoveries to allow the policy to learn how to respond when it departs from normal execution.

  3. Highlighting Task-Relevant Visual Cues: Integrate an adapted IAVLA framework where a large Vision-Language Model highlights task-relevant image regions via segmentation masks propagated across frames, providing task-relevant visual grounding rather than learning it from post-training data alone.

  4. Practical Tricks to Reduce VLA Dependence: Employ a Binary Hand Controller for manipulation tasks to control the hand as a gripper via a binary signal, and apply Butterworth Filtering to predicted action sequences at inference time to ensure executed actions are smooth and in-distribution.

  5. Experience-Driven Refinement Strategy: Adapt RECAP by conditioning the policy through a text-based instruction prefix (Advantage=True/False) and restricting the value function to camera observations and text instructions, allowing refinement without any architectural modification to the pretrained model.

  6. Value Function Design: Train a categorical distribution over return bins using only the Eagle-3 VisionLanguage encoder and a lightweight MLP head, ensuring the value function conditions only on the language instruction l and visual observation ot, not the robot state.

  7. In/Out-of-Distribution Analysis Tool: Implement a latent space analysis tool that fits a Gaussian Mixture Model to training embeddings and scores new states by their Mahalanobis distance to the nearest component, separating benign variation in joint configuration from genuine departures off the demonstrated manifold.

Sources

Related papers