Learning to Fold: prizewinning solution at LeHome Challenge 2026 (1st place online, 2nd offline)
# Learning to Fold: Bimanual Robot Places 2nd in Real-World Garment Challenge
A vision-language-action (VLA) policy won first place in simulation and second in real hardware at ICRA 2026's LeHome Challenge, a competition focused on bimanual robotic garment folding. The system combines a neural network that predicts both robot actions and task success probability into a single model, then refines its performance through reinforcement learning loops that learn from simulation and real-world trials alike.
The result demonstrates a practical approach to a coordination problem: getting two robot arms to manipulate deformable objects (garments) requires real-time visual feedback, semantic understanding of task progress, and decision-making under uncertainty. By unifying action prediction and success estimation in one network, the solution reduces the architectural complexity typical of multi-stage systems. The reinforcement learning component allows the policy to improve beyond its initial training rather than relying solely on imitation learning from human demonstrations.
The gap between first-place simulation performance and second-place real-world results reflects a persistent challenge in robotics: sim-to-real transfer. Even with reinforcement learning refinement, real sensors, actuator latency, fabric variability, and contact dynamics introduce failure modes not fully captured in simulation—a constraint that affects deployment timelines for automation integrators moving systems from development to production environments.