WALL-OSS Flow
WALL-OSS checkpoint with a Qwen2.5-VL backbone and flow-matching actions.
Research, evaluation and robot-specific training of WALL-OSS flow-matching policies.
Access and availability
- Code
- Public access
- Model files
- Public access
Training / fine-tuning code
- Access
- Public access
Verified on · 2026-10-30
- Code license
- Apache-2.0
Model files
| File | Access |
|---|---|
| x-square-robot/wall-oss-flow | Access |
Getting started
- Runtime and frameworks
- LeRobot / PyTorchNVIDIA GPU · bfloat16
- Fine-tuning requirements
- LeRobot integration: use the linked policy-specific setup and training instructions; compatibility with every upstream checkpoint is not implied.
At a glance
- Model type
- Vision-language-action model
- Backbone
- Qwen2.5-VL with flow-matching action head
- Weight parameters
- 4.224 B
- Input
- Language instruction / RGB camera images / Robot state / proprioception
- Output
- Continuous robot action chunks
Technical details
Continuous robot action chunks
Languages
English
Functional purpose
Action & control
Pipeline task
Robotics
Model card
Related paper: [Igniting VLMs toward the Embodied Space](https://arxiv.org/abs/2509.11766), arXiv:2509.11766, submitted 2025-09-15. The manuscript is dated 2025-09-08. Affiliation: X Square Robot. Correspondence: Hao Wang. Project lead: Lucy Liang. Andy Zhai appears first. The full author list is alphabetized by given name; the paper does not declare an ordering policy, so this does not establish contribution ranking. The paper introduces WALL-OSS for robot manipulation, combining vision-language understanding, instruction reasoning, subtask decomposition and action generation. This is the WALL-OSS family paper, distinct from the later WALL-OSS-0.5 paper.
Before you use this version: Requires task- and embodiment-specific data, configuration and validation. Published examples and simulated results do not establish safe operation on a physical robot. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.