💻 computer science

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

This paper introduces DEED, a systems-level framework that bridges the gap between lab benchmarks and real-world retail operations for humanoid robots by combining data-efficient post-training, experience-driven refinement, and latent-space analysis to achieve robust performance on a chip-restocking task using minimal computational resources.

Original authors: Roger Sala Sisó, Tiago Silvério, Jakob Sand, Tran Nguyen Le

Published 2026-07-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Roger Sala Sisó, Tiago Silvério, Jakob Sand, Tran Nguyen Le

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to do a chore, like folding laundry or restocking a shelf. You might think the hardest part is building the robot's brain so it can "see" and "understand" the world. But in the world of robotics, there's a tricky gap between what happens in a perfect, controlled laboratory and what happens in a messy, real-life store. In the lab, robots often look like geniuses, but step them out the door, and they might trip over a rug or drop an item because the lighting changed or a box was slightly crooked.

To bridge this gap, scientists use something called a Vision-Language-Action (VLA) model. Think of a VLA as a super-smart robot brain that has read the entire internet. It knows what a "chip bag" looks like, it understands the sentence "put this on the shelf," and it knows how to move its arms to do it. These models are powerful, but they are like students who have only ever studied from textbooks. They haven't actually done the homework in the real world. They struggle when things don't go exactly as planned, and they can't learn from their own mistakes once they are turned on. The big question researchers are asking is: How do we take these brilliant but inexperienced robot brains and turn them into reliable workers without needing a million dollars in supercomputers?

This paper introduces a new method called DEED (Data-Efficient Post-Training and Experience-Driven Learning) to solve exactly that problem. The researchers tested their idea on a Unitree G1-Edu robot, a humanoid that looks a bit like a human, trying to perform a very specific task: restocking bags of chips in a supermarket. They found that the secret to making the robot work wasn't inventing a new, more complex brain architecture. Instead, it was about being a very careful teacher.

First, they realized that simply downloading a pre-trained robot brain and telling it to "go" didn't work; the robot failed completely. The problem was that the robot was trying to learn from messy, inconsistent data and was confused by the timing of its cameras and movements. The team fixed this by creating a "Data-Efficient" recipe. They cleaned up the training videos to remove hesitations and mistakes, synchronized the robot's camera speed with its movement speed (so it didn't try to move faster than it could see), and even used a helper tool to highlight exactly where the robot should look, like drawing a green box around the empty spot on the shelf. They also simplified the robot's job by treating its hand like a simple on/off switch rather than trying to control every tiny muscle. With just these careful tweaks and a single graphics card, they turned a robot that was 0% successful into one that could restock chips 32% of the time.

Next, they tried to make the robot even better by letting it learn from its own experience, a process called "Experience-Driven Learning." They let the robot try the task on its own, and when it messed up, a human stepped in to fix it. The robot then used these corrections to update its brain. This worked! After one round of this practice, the robot's success rate jumped to 42%, and it became faster and more direct in its movements.

However, the paper suggests a cautionary tale about doing this too much. When they tried a second round of practice, the robot actually got worse, dropping back down to 22% success. The authors suspect this happened because the robot started relying too much on its own "hallucinations" of what worked, rather than the solid examples given by the human teacher. It's like if a student only practiced by guessing the answers to their own made-up tests; eventually, they start forgetting the real rules. The team also built a special tool to watch the robot's brain in real-time, measuring how "strange" its current situation was compared to what it had been trained on. This tool helped them see exactly when the robot was drifting into dangerous, unfamiliar territory.

Ultimately, the paper suggests that the biggest hurdle for putting humanoid robots in stores isn't building a smarter brain, but rather building a better training system. By carefully curating the data and knowing when to stop the robot from "practicing" on its own, we can turn a clumsy, pre-trained model into a competent worker using very little computing power. The researchers found that while one round of self-correction helps, too much can hurt, and the key to success lies in balancing the robot's own experiences with the reliable guidance of human demonstrations.

Technical Summary: Closing the Lab-to-Store Gap with DEED

Problem Statement
Despite the emergence of Vision-Language-Action (VLA) models as a dominant paradigm for general-purpose robotic manipulation, a significant gap remains between benchmark performance and reliable real-world deployment. Current VLAs are typically trained offline and deployed as fixed policies, lacking robustness against execution errors, distribution shifts, and environmental variability. The paper argues that bridging this "lab-to-store" gap is primarily a systems integration challenge rather than an architectural one. Deploying pretrained models (such as GR00T) often fails due to engineering mismatches in data curation, control frequency, and action representation, rather than a lack of model capacity. Furthermore, offline policies cannot improve from their own deployment experience, necessitating effective post-training strategies.

Methodology: The DEED Framework
The authors propose DEED (Data-Efficient Post-Training and Experience-Driven Learning), a systems-level framework evaluated on a supermarket chip-restocking task using a Unitree G1-Edu humanoid robot and the GR00T N1.6 foundation model. DEED consists of three core components:

Data-Efficient Post-Training Pipeline:
This stage focuses on converting an out-of-the-box foundation model checkpoint into a reliable policy through specific engineering decisions:
- Frequency Hierarchy: Establishing a strict relationship between camera frequency ( $f_{cam}$ ), teleoperation frequency ( $f_t$ ), recording frequency ( $f_r$ ), and inference control frequency ( $f_{ctrl}$ ) to ensure smooth, continuous motion and avoid learning from duplicated observations.
- Data Curation Rules: Implementing strict guidelines to prevent noise, including ensuring balanced state coverage, recording only efficient successful behaviors, retaining incidental recoveries (to teach error correction), maximizing action consistency, minimizing uncontrolled environmental variation, and avoiding "no-operation" states.
- Task-Relevant Visual Highlighting: Integrating an adapted IA-VLA framework where a large Vision-Language Model (VLM) generates segmentation masks to highlight task-relevant regions (e.g., empty shelf locations), providing the VLA with grounded visual cues rather than relying solely on post-training data.
- Reduced VLA Dependence: Utilizing a binary hand controller for discrete grasp actions and applying Butterworth filtering to predicted action sequences at inference to smooth transitions and keep executed actions within the training distribution.
- Continuous Episodes: Recording full multi-subtask demonstrations as continuous episodes to teach transitions between subtasks, rather than segmenting them.
Experience-Driven Learning (RECAP Adaptation):
To enable the policy to learn from its own deployment, the authors adapt the RECAP (RL with Experience and Corrections via Advantage-conditioned Policies) method to the decoupled GR00T architecture (separating perception/reasoning from control).
- Architecture Adaptation: Instead of modifying the model, the policy is conditioned via a text-based advantage prefix ("Advantage=True/False") and a vision-language-only value function.
- Value Function: A lightweight MLP head is trained on a frozen Eagle-3 VLM encoder to predict categorical return distributions over visual observations and instructions.
- Advantage Computation: Multi-step advantage labels are computed offline using Monte Carlo returns. Human corrective interventions are force-labeled as positive to reinforce recovery behaviors.
- Refinement Loop: The policy is iteratively refined by training on a mix of original teleoperation data, successful autonomous rollouts, and human corrections. Crucially, each refinement iteration is initialized from the original base checkpoint to prevent optimization drift.
Latent-Space In/Out-of-Distribution (OOD) Analysis:
A diagnostic tool is introduced to measure how far the robot's current state deviates from the training distribution. By fitting a Gaussian Mixture Model (GMM) to the latent embeddings of the state encoder, the system calculates Mahalanobis distances. This allows for the detection of states where the policy is likely to fail, distinguishing between benign variation and genuine departures from the demonstrated manifold.

Key Results
The framework was evaluated on a chip-restocking task involving grasping, torso rotation, and placement.

Baseline Failure: A naive fine-tuning of GR00T N1.6 (without DEED engineering) achieved a 0% success rate, failing to reliably grasp or place items.
Data-Efficient Stage: Applying the Data-Efficient recipe alone improved the success rate to 32% (16/50 episodes) with a mean execution time of 24.30s per bag. This was the first configuration to successfully execute the full pick-rotate-place sequence autonomously.
Experience-Driven Stage: A single RECAP iteration improved the success rate to 42% (21/50) and reduced execution time to 22.37s.
Degradation in Second Iteration: A second refinement iteration caused performance to drop to 22%. Analysis revealed this was due to distributional drift: self-generated rollouts began to dominate the training set, narrowing the effective state-action distribution and causing the policy to drift away from the robust behaviors anchored by the original teleoperation data.
Reset Behavior: Refined policies tended to stop after a single placement rather than returning to the start configuration, as the terminal success state was assigned a higher value than the intermediate reset poses.

Significance and Claims
The paper claims that the primary barrier to deploying VLA-based humanoids in real-world retail settings is not architectural novelty but systems integration. The authors demonstrate that:

Data Quality over Quantity: Careful data curation and targeted post-training can transform a non-functional foundation model checkpoint into a competent real-world system using only a single GPU.
Complementary Stages: The Data-Efficient stage provides the necessary foundation for functionality, while Experience-Driven refinement offers marginal gains in robustness.
Limits of Iterative Refinement: In high-variability humanoid tasks, iterative refinement without a balanced mixture of demonstrations and rollouts can lead to performance degradation due to distributional drift.
Monitoring as a First-Class Citizen: The proposed latent-space analysis tool provides a model-agnostic mechanism to diagnose deployment failures and understand the boundaries of a policy's training distribution.

The authors conclude that while DEED successfully bridges the gap for a specific task, the benefits of experience-driven learning are contingent on task variability and the maintenance of a balanced training distribution.

Technical Summary: Closing the Lab-to-Store Gap with DEED

More like this