lehome-challenge-submission

LeHome Challenge — Lakesenberg Submission

Submission package for the ICRA 2026 LeHome Challenge (bi-manual garment folding on the SO-101 platform). Generated on 2026-05-01.

---

⚡ TL;DR for evaluators (3 commands)

# 0. From inside an already-installed LeHome Challenge env
#    (uv venv + isaaclab + lehome package; see official docs/installation.md).
cd lehome-challenge

# 1. Pull this submission and install the dp_b1k plugin (10 s + 1 GB download).
git clone https://huggingface.co/Lakesenberg/lehome-challenge-submission submission_pkg
uv pip install -e ./submission_pkg/source/lerobot_policy_dp_b1k

# 2. Run our recommended policy (Solution F = dp_b1k) on top_long.
python -m scripts.eval \
  --policy_type lerobot \
  --policy_path submission_pkg/checkpoints/dp_b1k_four_types/pretrained_model \
  --dataset_root Datasets/example/top_long \
  --garment_type top_long \
  --num_episodes 12 \
  --enable_cameras --device cpu

Then loop --garment_type over top_long / top_short / pant_long / pant_short to obtain the four-category success rates.

💡 Use --policy_type lerobot for both submitted checkpoints. The custom dp_b1k policy is registered through LeRobot's standard third-party plugin mechanism (lerobot_policy_* namespace), so the eval script picks it up automatically once the plugin is pip install-ed.

---

📦 Repo contents

checkpoints/
  act_four_types/pretrained_model/         # Baseline (LeRobot ACT, 30k steps)
  dp_b1k_four_types/pretrained_model/      # Solution F (DP + B1K-style, 30k steps)
source/
  lerobot_policy_dp_b1k/                   # Custom LeRobot plugin (registers `dp_b1k`)
configs/                                   # Training configs (YAML)
rollout_results.txt                        # Local rollout summary (see env notice)
README.md                                  # This file

---

🤖 Submitted policies

Two policies are provided. Solution F (dp_b1k) is our primary entry; ACT is included as a baseline reference.

F. dp_b1k — DiffusionPolicy + BEHAVIOR-1K-style training/inference (PRIMARY)

Custom LeRobot plugin built on top of lerobot.policies.diffusion, with three additions ported from the BEHAVIOR-1K winning solution (originally Pi0.5/openpi):

• Correlated action noise — diffusion noise sampled with the empirical action-covariance Cholesky factor instead of i.i.d. Gaussian. Produces smoother action chunks that match the temporal statistics of human teleoperation.
• Soft inpainting at inference — the head of every action chunk is softly constrained to continue the previously executed actions, with covariance correction so the inpainted prefix doesn't break the noise structure. Removes the inter-chunk jitter that plain action-chunking suffers from.
• Optional depth branch — DiffusionDepthEncoder shares the RGB ResNet trunk with a separate single-channel stem; depth features are concatenated to the global conditioning vector. Disabled in the released checkpoint because the four_types_merged dataset does not ship observation.top_depth.

Plugin entry-point: lerobot_policy_dp_b1k.DpB1kPolicy, auto-registered as type dp_b1k via the lerobot_policy_ plugin convention.

A. ACT (baseline)

Stock LeRobot ACT trained on the same four_types_merged dataset (1000 episodes, 12-D bi-arm state/action, RGB only). Provided for reproducibility of the comparison reported below.

---

🧪 How to evaluate (full reproduction)

1. Set up the official LeHome Challenge environment

Follow the official docs/installation.md:

git clone https://github.com/lehome-official/lehome-challenge.git
cd lehome-challenge
uv sync
cd third_party && git clone https://github.com/lehome-official/IsaacLab.git && cd ..
source .venv/bin/activate
./third_party/IsaacLab/isaaclab.sh -i none
uv pip install -e ./source/lehome
# Download assets + example dataset (see docs/datasets.md)

2. Pull this submission and install the dp_b1k plugin

git clone https://huggingface.co/Lakesenberg/lehome-challenge-submission submission_pkg
uv pip install -e ./submission_pkg/source/lerobot_policy_dp_b1k

3. Evaluate both policies on all four categories

GARMENTS="top_long top_short pant_long pant_short"
POLICIES="act_four_types dp_b1k_four_types"

for G in $GARMENTS; do
  for P in $POLICIES; do
    python -m scripts.eval \
      --policy_type lerobot \
      --policy_path submission_pkg/checkpoints/${P}/pretrained_model \
      --dataset_root Datasets/example/${G} \
      --garment_type ${G} \
      --num_episodes 12 \
      --enable_cameras --device cpu
  done
done

--device cpu runs PhysX on CPU and renders cameras via GPU/Vulkan, which is the configuration recommended in docs/policy_eval.md.

Expected output

Each invocation prints per-episode success / failure logs and finishes with a Success Rate: X/12 line for the given category × policy combination.

---

📁 Local rollout note

rollout_results.txt reports N/A (env) for every category because the submitter's host (NVIDIA H200 + driver 550.163.01 cloud instance) cannot start Isaac Sim 5.1: the Vulkan/RTX render pipeline fails with VkResult: ERROR_DEVICE_LOST before any policy step is executed. We installed the full graphics stack (libnvidia-gl-550, libvulkan1, mesa-vulkan-drivers) and confirmed the H200 is visible to Vulkan 1.3.277, but the RTX ray-tracing pipeline initialization still fails on this driver version (a known mismatch fixed by NVIDIA driver ≥ 560 for Hopper GPUs).

What we did verify on the same host:

• Both checkpoints load cleanly into LeRobot 0.4.2.
• lerobot_policy_dp_b1k registers and scripts/smoke_test_dp_b1k.py passes both the forward and inference paths.
• Training completed to 30 000 steps for both ACT and dp_b1k.

Reproduction in the LeHome reference Isaac Lab 5.1 environment is expected to work normally.

---

🧩 Acknowledgments

• LeRobot — imitation-learning training stack (ACT and Diffusion Policy)
• Isaac Lab / Isaac Sim — simulation environment
• BEHAVIOR-1K winning solution (Pi0.5) — for the correlated-noise / soft-inpainting recipe that motivates Solution F