feat: support wallx model

[wut19]

This pull request introduces support for the new Wall-X policy, a Vision-Language-Action model for cross-embodiment robotic control, into the lerobot codebase. The changes include adding the Wall-X policy configuration, model, processor, and associated dependencies, as well as updating the policy factory to recognize and instantiate Wall-X. These updates enable users to train and deploy the Wall-X policy using the same unified interface as other policies in the library.

Wall-X Policy Integration

• Added the Wall-X policy implementation, including WallXConfig, WallXPolicy, and processor logic in the new src/lerobot/policies/wall_x/ module. This includes configuration validation, optimizer/scheduler presets, and pre/post-processing pipelines tailored for Wall-X. [1] [2] [3] [4]
• Added a README for Wall-X describing its capabilities, architecture, and citation information.

Policy Factory and Registry Updates

• Registered Wall-X in the policy factory, enabling instantiation and configuration via the standard API (get_policy_class, make_policy_config, and make_pre_post_processors). Factory docstrings updated to reflect Wall-X as a supported policy type. [1] [2] [3] [4] [5] [6] [7]
• Updated the policy module __init__.py to include WallXConfig in the export list for easier imports and discoverability. [1] [2]

Dependency Management

• Added a new optional dependency group wallx in pyproject.toml with required packages for training and inference (torch, transformers, accelerate, etc.), and included it in the all dependency group for convenience. [1] [2]

[Copilot]

Pull request overview

This pull request adds comprehensive support for the Wall-X policy, a Vision-Language-Action model for cross-embodiment robotic control based on Qwen2.5-VL with flow matching. The implementation integrates Wall-X into the LeRobot framework, providing a unified interface for training and deploying this advanced multimodal policy.

Key Changes:

• Implemented Wall-X policy with Qwen2.5-VL backbone and action prediction via flow matching
• Added configuration, processor, and model components with MoE architecture support
• Integrated Wall-X into policy factory system for seamless instantiation
• Added optional dependency group for Wall-X specific requirements

Reviewed changes

Copilot reviewed 13 out of 14 changed files in this pull request and generated 34 comments.

Show a summary per file

File	Description
tests/policies/wall_x/test_wallx.py	Test file for Wall-X policy integration (CI-skipped for local testing only)
tests/policies/wall_x/init.py	Test module initialization
src/lerobot/policies/wall_x/utils.py	Utility functions for data processing, text formatting, and action tokenization
src/lerobot/policies/wall_x/qwen_model/qwen2_5_vl_moe.py	Extended Qwen2.5-VL model with MoE architecture and action processing
src/lerobot/policies/wall_x/qwen_model/configuration_qwen2_5_vl.py	Configuration classes for Qwen2.5-VL vision and language components
src/lerobot/policies/wall_x/processor_wall_x.py	Pre and post-processing pipelines for Wall-X inputs/outputs
src/lerobot/policies/wall_x/modeling_wall_x.py	Main Wall-X policy implementation with action head and training logic
src/lerobot/policies/wall_x/constant.py	Constants for camera mappings, resolution, and preprocessing parameters
src/lerobot/policies/wall_x/configuration_wall_x.py	Wall-X configuration with validation and optimizer presets
src/lerobot/policies/wall_x/init.py	Module exports for Wall-X components
src/lerobot/policies/wall_x/README.md	Documentation describing Wall-X capabilities and citation info
src/lerobot/policies/factory.py	Factory updates to register and instantiate Wall-X policy
src/lerobot/policies/init.py	Added WallXConfig to policy module exports
pyproject.toml	Added wallx optional dependency group with required packages

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[pkooij]

Looks good overall, one TODO for me is test it on real robot (SO101), could you make flash-attention optional? Then once I know the correct torchcodec version I can kick off training and verify it on real robot.

For later: we do need to port preprocessing to pipeline steps