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

2026-06-25 · Source: Artificial Intelligence · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

The LeHome Challenge 2026 prizewinning solution, which secured 1st place among 62 teams in the online simulation round and 2nd in the real-world final at ICRA 2026, details a bimanual garment folding system. This system enhances a vision-language-action (VLA) policy through a reinforcement learning loop. Uniquely, the policy network functions as its own value function, predicting success, progress, and future task-relevant quantities. These predictions are crucial for driving advantage estimation, live failure detection, and candidate selection. The solution integrates existing RL concepts with engineering and optimization contributions, including AWR + RECAP for flow-matching VLA, an asynchronous distributed training/rollout pipeline utilizing HuggingFace Hub, inference-time hyperparameter optimization via Thompson sampling, and a robust sim-to-real recipe featuring camera-alignment tooling, extensive augmentation, and DAgger-like human-in-the-loop data collection.

Key takeaway

For Robotics Engineers developing bimanual manipulation systems for deformable objects, this LeHome Challenge solution provides a validated recipe. You should consider integrating its VLA policy, which doubles as a value function for success prediction, into your RL framework. Adopting its asynchronous distributed training via HuggingFace Hub and its robust sim-to-real pipeline, including DAgger-like HIL, can significantly accelerate your development and improve real-world performance.

Key insights

A bimanual garment folding system leverages a VLA policy with an RL loop, where the policy network also serves as its own value function.

Principles

• Policy networks can serve as their own value functions.
• AWR and RECAP can be combined for VLA flow-matching.
• Asynchronous distributed training enhances RL scalability.

Method

Improve a VLA policy via an RL loop where the policy network predicts success and progress for advantage estimation. Integrate AWR+RECAP, asynchronous distributed training, Thompson sampling for inference-time optimization, and a sim-to-real pipeline with DAgger-like HIL.

In practice

• Utilize HuggingFace Hub for distributed training.
• Apply Thompson sampling for inference hyperparameter optimization.
• Employ DAgger-like HIL for robust sim-to-real transfer.

Topics

• Bimanual Garment Folding
• Vision-Language-Action Policies
• Reinforcement Learning
• Sim-to-Real Transfer
• Distributed Training
• Thompson Sampling

Best for: MLOps Engineer, AI Engineer, Computer Vision Engineer, Robotics Engineer, Machine Learning Engineer, AI Scientist

Related on AIssential

• ReCoVLA: VLM-Guided Reward Compilation for Failure Recovery in Vision-Language-Action Policies — ReCoVLA uses a VLM to compile structured rewards for residual policy training, enabling robust VLA policy recovery.
• We Hired 10 Humans to Train a Robot to Fold Shirts — Robots can learn complex manipulation tasks like t-shirt folding through teleoperated data collection.
• LadderMan: Learning Humanoid Perceptive Ladder Climbing — LadderMan enables robust humanoid ladder climbing and manipulation through perceptive, sim-to-real learning.
• DuoBench: A Reproducible Benchmark for Bimanual Manipulation in Simulation and the Real World — DuoBench offers a reproducible benchmark for bimanual robot manipulation, revealing current policy limitations in coordination and sim-to-real transfer.
• Intelligent Robots in 2026: Are We There Yet? [Nikita Rudin] - 760 — Bridging the sim-to-real gap and integrating semantics are critical for advanced robot autonomy and value generation.
• DAM-VLA: Decoupled Asynchronous Multimodal Vision Language Action model — Decoupling temporal processing for vision, language, and action modalities significantly enhances VLA model performance and control robustness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.