Learning to Fold: prizewinning solution at LeHome Challenge 2026 (1st place online, 2nd offline)

A Tactile Triumph: What the LeHome 2026 Win Reveals About VLA Progress

The LeHome Challenge 2026, held at ICRA, tasked teams with a deceptively simple problem: getting two robot arms to fold a garment. The winning solution, which secured 1st place in simulation and 2nd in the real-world finals, is notable not just for its rank, but for what it signals about the maturation of vision-language-action (VLA) models in robotic manipulation.

The Technical Achievement

The system improves upon a standard VLA architecture, which processes visual input and language instructions to generate motor commands. The key innovation appears to be in how the model handles the sequential, deformable nature of fabric. Unlike rigid objects, garments have infinite degrees of freedom, making state estimation notoriously difficult. The winning approach likely integrates a learned dynamics model or a specialized perception pipeline that tracks keypoints on the fabric—such as corners and edges—even as they become occluded during folding.

The gap between the 1st place simulation result and the 2nd place real-world finish is itself instructive. Simulation provides perfect state information and repeatable physics, while reality introduces sensor noise, friction variability, and the "sim-to-real" transfer problem. That the system performed competitively in both domains suggests the VLA backbone is sufficiently robust, but the drop in ranking highlights that tactile feedback and real-time adaptation remain hard to simulate.

Why This Matters for AI Practitioners

First, deformable object manipulation is a benchmark for generalizable robotics. If a VLA can fold a shirt, it can likely handle other non-rigid tasks like bagging groceries or assembling cables. This competition demonstrates that the field is moving from "pick and place" to "pick, manipulate, and conform."

Second, the VLA paradigm is proving its worth in low-data, high-variation settings. Traditional robotic control requires thousands of hand-coded trajectories or extensive reinforcement learning. A VLA can leverage pre-trained language and vision embeddings to generalize from fewer demonstrations. For practitioners, this means that fine-tuning a large pre-trained model on a specific task like folding may be more efficient than building a bespoke controller from scratch.

Third, the simulation-to-real gap remains the critical bottleneck. The second-place real-world finish suggests that even the best VLA models struggle with the "last centimeter" of precision when physics becomes messy. Practitioners should invest in domain randomization during simulation training and consider adding a low-level feedback controller (e.g., impedance control) to correct for model errors in deployment.

Implications for the Road Ahead

This result reinforces that VLA models are becoming the default architecture for complex manipulation, but they are not yet a plug-and-play solution. The most successful systems will likely be hybrids: a high-level VLA for task planning and coarse motion, paired with a classical control loop for fine-grained adaptation. For AI researchers, the LeHome Challenge provides a concrete benchmark to measure progress—and a reminder that folding a shirt is harder than it looks.

Key Takeaways

• The winning solution demonstrates that VLA models can handle deformable object manipulation at a competitive level, bridging simulation and real-world performance.
• The gap between 1st in simulation and 2nd in reality underscores the persistent challenge of sim-to-real transfer, especially for tasks requiring precise contact and adaptation.
• Practitioners should adopt a hybrid architecture: a pre-trained VLA for high-level reasoning, supplemented by a classical feedback controller for low-level precision.
• Deformable object manipulation is emerging as a key testbed for generalizable robotic intelligence, with implications for logistics, healthcare, and domestic automation.