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

TL;DR

A single-researcher engineering case study of a flow-matching VLA (SigLIP + Gemma-2B prefix + Gemma-300M action expert from the BEHAVIOR-1K 2025 winner) improved with an RL loop that placed 1st of 62 teams in the LeHome Challenge 2026 simulation round and 2nd in the ICRA real-world final on bimanual garment folding. The recipe combines AWR (as a sampler re-weight rather than a loss weight) with RECAP-style advantage-conditioning for classifier-free guidance at inference, all on top of an asynchronous training/rollout pipeline that talks only through HuggingFace Hub. A single learned query token drives cheap linear heads for success, task completion, garment type, current and 30-frame-future keypoint distances, and an action-conditional success residual — making the policy its own value function, Q-function, and cheap world-model substitute. Also introduces Thompson-sampling bandits over inference-time hyperparameters (execution length, playback speed, inpainting onset, guidance scale, noise temperature, best-of-N).

Key claims

Reweighting/conditioning methods (AWR + RECAP) fit flow-matching VLAs better than PPO-family policy-gradient methods because valid actions occupy a thin manifold and log-probability-pushing objectives leave that manifold, whereas conditioning and reweighting only redistribute mass toward good actions the policy already produces [§2.2].
AWR is more compute-efficient when applied through sampling rather than through a loss weight: high-advantage frames are loaded more often, batches stay unweighted MSE, and effective batch utilization stays at 100% of the weight mass [§2.3].
Combining AWR with RECAP-style advantage conditioning is complementary — AWR reweights the mixture toward the “good actions” mode, RECAP conditioning selects the positive-advantage slice — with a toy-picture argument that doing both simultaneously leaves the target almost entirely on the good mode [§2.2, Fig. 3].
A single learned query token in the image group of the prefix, with cheap linear heads reading image tokens only, is enough to make the policy predict success, task completion, garment type, current and 30-frame-future keypoint distances, and an action-conditional success residual [§4.1, §5].
The action-conditional success residual head acts as a Q-function inside the same policy network, and predicted keypoint futures act as a cheap world-model substitute — all sharing representation with the action head and served with one forward pass [§5].
Per-frame advantage is built by combining a CUPED-style dampened value baseline (from the success head) with a completion-progress signal (from the completion head) via GAE, and degrades gracefully to outcome-only baselines as rollouts go stale [§6, §6.5].
Sparse binary success can be densified into per-garment checkpoints from the challenge’s own keypoint conditions with all reward withdrawn on failure so the episode return stays binary — engineered dense intermediate reward without changing the outcome distribution [§6].
Checkpoint rollback (roll back to a checkpoint from a few days ago and continue training on all data collected since, including data from newer checkpoints) reliably kicks the policy out of local optima that continuous training settled into; 3 rollbacks in round 1 and 1 in round 2 [§2.5].
An asynchronous three-component loop (training worker, N rollout workers, manual DAgger station) that communicates only through HuggingFace Hub scales data collection by “starting another machine” and never blocks the trainer on network or worker state [§2.1].
Inference-time hyperparameters (execution length, playback speed, inpainting onset, CFG scale, noise temperature, best-of-N candidate count) are per-garment-type and can be tuned online with a Thompson-sampling bandit during rollout collection — one checkpoint, many effective behaviors [§7].
The policy achieved 79.63% overall success on the simulation leaderboard, 6.1 points ahead of second place across 62 teams [§8]; and placed 2nd on the real-robot final with 865/1080 points [§9].
Sim-to-real (one-week sprint from a late-but-not-latest sim checkpoint) relied on heavy augmentation (episode-level texture/pose/lighting + per-step color/lighting), camera-alignment tooling, real-robot manual DAgger, and a mix of organizer BC data, teleop/DAgger, and augmented sim replays [§9].
Manual sim-side DAgger turned out to be of limited use because teleoperating through a sim interface was harder than the policy’s own folding by mid-training; real-robot DAgger, in contrast, was one of the project’s most useful tools [§3.4, §9.10].

Method

The system is a flow-matching VLA built on the author’s BEHAVIOR-1K 2025 winner: frozen SigLIP-So400m/14 image encoder, Gemma-2B prefix transformer over (images + query + state + garment-type token + advantage token + FAST tokens), and a Gemma-300M action expert that emits 30-step, 12-dim joint-delta action chunks (1 s at 30 Hz) by flow matching over the prefix KV cache. Three RGB cameras (top + wrist × 2) at 224×224; text input dropped entirely; correlated flow-matching noise + correlation-aware soft inpainting at chunk boundaries + cross-layer KV-cache mixing carried over from the BEHAVIOR-1K solution.

The LeHome additions are: a learned garment-type input token with a System-2-like inference-time bootstrap (predict once at episode start, use as input for the rest); RECAP-style advantage conditioning via a dedicated advantage token gated by stochastic masking (25→10% unconditional-dropout rate ramp), unlocking CFG at inference; AdaRMS multi-signal conditioning that feeds the advantage vector into every action-expert layer’s RMSNorm; exclusive self-attention; smooth per-timestep action normalization; and a set of cheap linear heads reading only from the image tokens (success, completion, garment type, current keypoint distances, 30-frame-future keypoint distances, action-conditional success residual).

RL is an asynchronous three-way loop over HuggingFace Hub: a training worker (single H200, batch 192, ~300k steps, cosine LR $10^{-4}$ → $10^{-6}$ , AdamW, bf16, frozen SigLIP) recomputes advantages across all rollout datasets every iteration and uploads checkpoints every 500 steps; N rollout workers (mostly RTX PRO 6000 GPUs) run 3–5 parallel Isaac Sim instances behind one stateless policy server, sampling episodes via multiple strategies (random, full sweep, difficulty-targeted curriculum after [14], success-state replay with heavy augmentation, semi-success replay handoff, hard-mining failure-state replays); a human DAgger station teleoperates from saved failure/semi-success states and uploads corrections. Advantage is consumed both through the sampler (AWR: high-|advantage| frames loaded more often, importance-weighting the auxiliary-head losses back to unbiased) and through the network (RECAP: advantage in as conditioning, driving CFG at inference).

Reward densification uses the challenge’s own keypoint conditions as per-garment checkpoints, with all reward withdrawn on failure so the episode return stays binary; the success head supplies a CUPED-dampened value baseline, the completion head supplies progress, and GAE combines them into per-frame advantages. Inference-time hyperparameters are per-garment-type and tuned online by a Thompson-sampling bandit during rollout collection.

Results

Simulation round (Feb–Apr 2026): 1st of 62 teams, 79.63% overall success across four garment types × 20 instances (10 seen, 2 public unseen, 8 private unseen); 6.1-point margin over 2nd place. Per-type: 74.5% long tops / 70.0% short tops / 80.5% long pants / 93.5% short pants — top score outright on short tops, long pants, short pants; 3rd on long tops (behind 77.0% and 76.5%) [§8].
Real-world final (June 2026, ICRA Vienna): 2nd of 8 finalists, 865/1080 points on the organizers’ combined success + fold-quality metric; unseen garments (50% score weight) folded on a bimanual SO-ARM101 rig the author never had access to before the event [§9].
Recipe was a one-person effort under competition pressure, “an engineering case study, not a controlled experiment” [abstract] — no ablations of the individual pieces; recipe pieces released as one bundle plus checkpoint (huggingface.co/IliaLarchenko/lehome_sim).

Why it’s interesting

Sharpest filed evidence that the π*0.6: a VLA That Learns From Experience (RECAP) RECAP recipe (RL-fine-tuning flow-matching VLAs via advantage conditioning + CFG) can be reproduced by a single researcher on a tight budget and reach top-of-leaderboard status on a real ICRA competition — the tightest independent replication yet, tighter than the Evo-RL: Open Real-World Offline RL on SO-101 and AgileX PiPER Evo-RL SO-101 port and complementary to it (Evo-RL ships the pipeline; this paper adds a full self-value-function head-stack, a Thompson-sampling inference tuner, and a manual DAgger loop). Distinct move on top of RECAP: instead of a separate value model (as in World Value Models for Robotic Manipulation) or a binarized text-token value (π*0.6: a VLA That Learns From Experience (RECAP) §IV-B), the policy is its own value function, Q-function, and cheap world-model substitute, all driven by one learned query token — a structural counterpoint on the “value-head vs value-model” axis the VLA Models concept tracks. AWR-through-the-sampler is a compute-efficiency insight orthogonal to the advantage-conditioning literature and worth pulling on for the wiki’s Reasoning RL line.