Wall-OSS-0.5 for LeRobot

This repository contains the LeRobot-format conversion of x-square-robot/wall-oss-0.5 at revision b7dc23a82dbf70a65859c9fdd596703234597028.

The policy uses the audited X-Square-Robot/wall-x runtime at commit 6764e8f12f320819e86d5476b3ed68fb8f45b43c. The original 1,061 tensor keys, tokenizer, processor metadata, continuous-flow normalizer, proprioception normalizer, and autoregressive normalizer are preserved.

Validation

Author-versus-LeRobot parity was evaluated in BF16 on an RTX 5090 for batch sizes 1 and 2. Prompt tokens, attention masks, transformed images, normalized state/actions, flow velocity and loss, 10-step Euler samples, and final denormalized actions were exact (max_abs_error = 0).

A real-checkpoint fixed-batch training smoke reduced loss from 0.377876 to 0.105791. Saved and reloaded weights and loss were exact, and a resumed optimizer update produced finite nonzero gradients.

No Wall-OSS-0.5 RoboCasa result is reported: neither the official author organization nor the conversion machine contains a post-trained Wall-OSS-0.5 RoboCasa checkpoint.

Usage

The LeRobot policy integration is proposed in huggingface/lerobot#4200.

git clone https://github.com/X-Square-Robot/wall-x.git
git -C wall-x checkout 6764e8f12f320819e86d5476b3ed68fb8f45b43c
export PYTHONPATH="$PWD/wall-x${PYTHONPATH:+:$PYTHONPATH}"

lerobot-train \
  --policy.type=wall_oss_05 \
  --policy.pretrained_name_or_path=lerobot/wall-oss-0.5 \
  --dataset.repo_id=your-org/your-dataset

Wall-OSS-0.5 uses a Qwen2.5-VL backbone, but the released runtime adds custom MoE routing, proprioception/action embeddings, and continuous-flow action generation. It is therefore not a plain Transformers Qwen2_5_VLForConditionalGeneration checkpoint.