What Is WALL-OSS?

WALL-OSS is introduced in the paper Igniting VLMs toward the Embodied Space (arXiv 2509.11766) from the X Square Robot team, dated 8 September 2025 in the paper text. The authors argue that existing vision-language models still have limited spatial and embodiment understanding, and that action comprehension and generation is the central bottleneck. The paper, the official GitHub repository and the Hugging Face model page all use the name WALL-OSS.

WALL-OSS is third-party research. Firsthand is not part of it or affiliated with its authors, and makes no claim that any model was trained on Firsthand data. What the site states is custom egocentric collection of real human demonstrations, fast and on demand, to a written spec: 3D hand pose at 21 joints per hand, calibrated stereo depth and 6-DoF IMUs at 200 Hz, delivered in RLDS, LeRobot format or WebDataset with a buyer-owned commercial license. The only language data the site states is one task instruction per episode.

The paper names three gaps in moving a vision-language model to action: modalities and data scale, pretraining distributions, and training objectives. On distributions it says embodied data typically involves first-person viewpoints, fisheye imaging and self-occlusions, which differ substantially from internet imagery.

On objectives it says language models use next-token likelihood on discrete sequences, whereas action trajectories are continuous, high-frequency signals.

The paper adopts Qwen2.5-VL-3B as the main backbone. The model takes vision, described as egocentric and arm-mounted camera views, together with text instructions. It uses a tightly coupled Mixture-of-Experts design in which a static router sends action features to an Action feed-forward network and vision-language features to a Vision-Language feed-forward network, instead of a learned router.

In the Inspiration stage the vision-language model is trained with embodied question answering and discrete action tokens produced by FAST tokenization, giving coarse action awareness. In the Integration stage discrete action prediction is replaced by continuous action modeling with flow matching.

The Integration stage has two phases: first the vision-language model is frozen and only the flow head is trained, then the model is unfrozen for joint optimization.

The paper generalizes chain-of-thought to span instruction, reasoning, subtask plan and continuous actions within a single model. At inference the model can use the intermediate steps or map an instruction directly to actions, and can interleave reasoning with execution.

The paper describes three components: self-collected robot action data, open-source action data, and multimodal question-answering data. Named open-source action sources include DROID, BC-Z and BRIDGE. For its self-collected data the authors describe step annotations with human spot checks, multi-sensor timestamp synchronization, outlier filtering and removal of idle frames.

The paper and the project page give different phrasings of the corpus size, so this entry quotes no hour count.

The authors evaluate an embodied question-answering benchmark and six manipulation tasks, with blind third-party evaluators, and compare against pi-0 and Diffusion-Policy baselines. In the paper’s own summary, pi-0 remains superior for precise manipulation.

These are the authors’ claims on their own setups. Firsthand does not reproduce them. The one comparison reproduced here is the embodied question-answering table below.

A vision-language model, as the site defines it, reads images and text. WALL-OSS starts from one and adds action generation, so the same model reasons about a scene and emits actions. It is an instance of the robot foundation model and vision-language-action model ideas the site defines, and it uses the FAST tokenization the site covers. Firsthand ranks none of these.

This entry draws on the arXiv paper, the official project page, the official GitHub repository and the Hugging Face model page. The GitHub repository states the Apache License 2.0 for the code. The licence of the model weights is Not stated here. The sources do not state that human egocentric video, stereo depth, hand pose or IMU data was used, so how those would suit WALL-OSS is Not stated.

When a program needs real human egocentric demonstrations of tasks, environments or objects that public data does not cover, specified in writing. Firsthand does not build or train models, and the language it delivers is one task instruction per episode.

Confirm the spec on a Spec Pilot, a small fast first batch. Where coverage is deep, the median from signed spec to first validated batch is 48 hours; where it is thin, three to four weeks. Billing is per validated hour.

What is WALL-OSS?

A published open-source embodied foundation model from X Square Robot (arXiv 2509.11766) that adapts a vision-language model to robot control in one end-to-end model.

Which backbone does WALL-OSS use?

The paper adopts Qwen2.5-VL-3B as the main backbone, extended with action experts in a Mixture-of-Experts design.

What is Unified Cross-Level chain-of-thought?

The authors’ term for one model spanning instruction reasoning, subgoal decomposition and fine-grained action synthesis, with intermediate steps optional at inference.

Is the code or model open?

The GitHub repository provides training and inference code under the Apache License 2.0 and lists checkpoints on Hugging Face. The weights licence is Not stated in the sources read.

Does Firsthand data train WALL-OSS?

No claim is made. Firsthand is not affiliated with its authors, and the site does not state that any model was trained on Firsthand data.

Can Firsthand collect egocentric data for a VLA program?

The site states custom egocentric collection to a written spec, fast and on demand, with 21-joint 3D hand pose, calibrated stereo depth, 200 Hz IMUs and one task instruction per episode, confirmed on a Spec Pilot.