Wall-OSS-0.5 for LeRobot

This repository contains the LeRobot-format conversion of x-square-robot/wall-oss-0.5. The policy is implemented natively in LeRobot with PyTorch and Transformers; no separate Wall-X runtime checkout is required.

The checkpoint contains the 4B model weights, Qwen processor and tokenizer, LeRobot policy configuration, and saved LeRobot pre/postprocessors. State and action q01/q99 statistics live in the processors. When fine-tuning with lerobot-train, the target dataset statistics replace these pretrained values and are saved with the resulting checkpoint.

The base model uses canonical 26-dimensional state and action features, 32-step action chunks, and a 10-step continuous-flow sampler. Fine-tune it on the target robot and task dataset before deployment.

Usage

The LeRobot policy integration and documentation are available in huggingface/lerobot#4200.

pip install "lerobot[wall_oss_05]"

lerobot-train \
  --policy.path=lerobot/wall-oss-0.5 \
  --dataset.repo_id=your-org/your-dataset

Wall-OSS-0.5 uses Transformers for its Qwen2.5-VL components. LeRobot adds the Wall-specific routed action experts, state serialization, flow objective, and Euler sampler, so it is not a plain Transformers Qwen2_5_VLForConditionalGeneration checkpoint.

Native checkpoint layout (2026-10-22)

The weights now use the native LeRobot policy state dictionary, with standard shared loading and distributed saving. All 837 neural tensors are preserved exactly from the earlier conversion. Normalization remains processor-owned; the current split runtime/training renderers and tokenizer are serialized. The config uses a stable tokenizer asset reference and can load from another working directory. Use --policy.path=lerobot/wall-oss-0.5 to load the weights.

The policy supports native state/action features up to 26D when fine-tuned with matching dataset statistics. Padded joints are masked and output actions use the action feature width independently of state width. Image encoding and prompt prefill run once per action chunk; all ten flow steps reuse the prefix.

These changes accompany LeRobot PR #4200, head 37df61eab95b2dcc030d9c01577346106f6c3a42. Until it is merged, use that PR's code. The previous layout remains available at revision eb780f31d4bd3e5abb87fdd102435a4d68e374d5 for older PR revisions.