Open Source RoboticsRobots, software & learning

Wall-X

Robotics model

VLA training/inference source for WALL-series robot policies, with separate published checkpoint listings and CUDA operator builds. [1] [3]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Robot-learning researchers working with CUDA and LeRobot data [1]
Parameters
4,162,224,832 (~4.16B [31]
Embodiments
Franka (RoboChallenge / DROID), UR5 (Berkeley Autolab / RoboChallenge), ALOHA (RoboChallenge), … [31]
Source rights
Some licenses are unresolved. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
Ecosystem
LeRobot
License
Apache-2.0 code; checkpoint rights separate · license not fully assessed for model weights · Details by part

Requirements · Research details and dates

On this page

Overview

Wall-X (x-square-robot/wall-x) provides the training, fine-tuning, and serving codebase for the WALL-series Vision-Language-Action (VLA) models, paired with the public x-square-robot/wall-oss-0.5 checkpoint on Hugging Face. At pinned Hugging Face revision b7dc23a82dbf70a65859c9fdd596703234597028, wall-oss-0.5 contains 4,162,224,832 parameters (~4.16B: 4,158,769,152 BF16 + 3,455,680 F32) in an 8.33 GB ungated model.safetensors file. It builds on a 3B Qwen2.5-VL backbone (Qwen2_5_VLForConditionalGeneration, 36 layers, hidden_size: 2048) augmented with action-generation modules (action_hidden_size: 1024, state_hidden_size: 2048) trained via gradient-bridged co-training across discrete action prediction, multimodal prediction, and continuous flow matching. [1] [3] [30] [31]

Model card summary

Architecture

Qwen2.5-VL 3B vision-language backbone (Qwen2_5_VLForConditionalGeneration, 36 layers, hidden_size 2048) augmented with dedicated action-generation modules (action_hidden_size 1024, state_hidden_size 2048) trained with gradient-bridged co-training across discrete action prediction, multimodal prediction, and continuous flow matching.

[31]

Base model

Qwen2.5-VL (3B VLM backbone; Qwen2_5_VLForConditionalGeneration)

[31]

Parameter count

4,162,224,832 (~4.16B; 4,158,769,152 BF16 + 3,455,680 F32 per wall-oss-0.5 safetensors metadata)

[31]

Input modalities
rgbproprioceptionlanguage

Multi-view RGB camera streams (e.g., face_view and right_wrist_view at 448px in config.yml), robot proprioceptive state (512 bins / string representation), and language instructions.

[1] [31]

Output modalities
ee-pose-deltagripper-command

Default wall-oss-0.5 config.yml defines 26-D relative dual-arm EE Cartesian position + 6D rotation + grippers plus decomposed base velocity, height, and head actions; configurable via embodiment norm_key.

[31]

Action space

Continuous flow-matching and autoregressive discrete action chunks (action_horizon_flow: 32, action_horizon_ar: 32, state_bins: 512); default wall-oss-0.5 config.yml specifies 26-D relative dual-arm EE pose, grippers, velocity, height, and head actions with per-embodiment norm_key normalization.

[31]

Trained / supported embodiments
  • Franka (RoboChallenge / DROID)
  • UR5 (Berkeley Autolab / RoboChallenge)
  • ALOHA (RoboChallenge)
  • ARX5 (RoboChallenge)
  • 20+ pretrained robot embodiments via norm_key (BridgeData v2, RH20T, AgiBot World Alpha, DobbE, UMI-biarm, Galaxea LeRobot v2.1, RoboCoin)

[31]

Training datasets
  • Multi-embodiment mixture (>1M trajectories/epoch across 20+ embodiments including DROID, BridgeData v2, RH20T, AgiBot World Alpha, Fractal, FurnitureBench, BC-Z, RoboChallenge, RoboCoin) plus grounded multimodal corpus and proprietary/internal configurations (ex_normal, x2_normal)

[31]

Interfaces
  • Wall-X Python CLI/serving utilities
  • Hugging Face Hub safetensors
  • LeRobot dataset format

[1] [31]

Checkpoints

CheckpointFormatSizeAccessPinned revision
x-square-robot/wall-oss-0.5safetensors8.33 GBopenb7dc23a82dbf

Ungated 8.33 GB model.safetensors verified via Hugging Face API at revision b7dc23a82dbf70a65859c9fdd596703234597028; no license field is declared in the model card metadata.

[30] [31]

Is this for you?

Use cases
  • Zero-shot and fine-tuned multi-embodiment manipulation policy evaluation
  • VLA fine-tuning on LeRobot-formatted robot demonstration datasets

[1] [31]

Known limitations
  • WALL-OSS0.5 card and repository Python environment instructions differ.
  • Selected checkpoint is public/ungated in metadata, but no checkpoint license declaration was identified in card/metadata; code Apache2 does not settle weight rights.
  • No CUDA build, checkpoint payload, inference, closed-loop control or data rights validated.

Editorial scope assessment from the cited implementation/documentation; no execution or compatibility test.

[1] [2] [3] [4] [30] · Editorial assessment

Safety and limitations
  • WALL-OSS-0.5 model card (Python 3.12) and wall-x repository README (Python 3.10) specify different Python environment routes that must be reconciled before installation.
  • The wall-oss-0.5 checkpoint is public/ungated, but no weight license declaration was identified in card/metadata; repository Apache-2.0 does not automatically settle weight or training-data rights.
  • Pretraining mixture includes proprietary/internal configurations (ex_normal, x2_normal) alongside public benchmarks.
  • No CUDA build, weight execution, inference, or physical closed-loop robot control was tested by Open Source Robotics.

[1] [2] [30] [31]

Compute & fine-tuning

Inference compute
  • Documented hardware: NVIDIA CUDA GPU (flash-attn 2.8.3 and compiled CUDA operators required)
  • Precision: bf16
  • Edge / consumer capable: No

8.33 GB BF16/F32 safetensors checkpoint with FlashAttention-2 and custom CUDA operators; exact minimum VRAM and latency_ms are not reported.

[1] [31] · Editorial assessment

Fine-tuning

[1] [31]

Host computer

CUDA-capable NVIDIA GPU required for custom CUDA operators and flash-attn 2.8.3; exact minimum VRAM is not reported in upstream documentation.

[1] [31]

Operating system
  • Linux

Inferred from CUDA operator compilation and flash-attn 2.8.3 build instructions.

[1] [31] · Editorial assessment

Dependencies
  • Repository route uses Python3.10 while selected WALL-OSS0.5 model card uses Python3.12
  • Requirements, flash-attn2.8.3, DMuon and separately installed LeRobot are documented
  • CUDA operator compilation during editable install requires a matching PyTorch/CUDA build environment

[1] [2] [3] [4] [30]

Setup & usage

Open the upstream getting-started guide [1]

  • Select WALL-OSS0.5 or another named checkpoint and its matching norm_key/action setup; reconcile model-card and code-repository environment instructions before attempting installation.
  • Preserve exact source/dependency/model versions and evaluate the selected workflow before connecting physical hardware or reusing data.

Editorial scope assessment from the cited implementation/documentation; no execution or compatibility test.

[1] [2] [3] [4] [30] · Editorial assessment

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
Wall-OSS-0.5 zero-shot real-robot suite (17 tasks, incl. held-out deformable manipulation)Upstream model card claims non-trivial zero-shot completion directly from the pretrained checkpointUpstream-reported claim on the Hugging Face model card; not independently executed or verified by OSR.
Wall-OSS-0.5 fine-tuned real-robot suite (15 tasks)60.5% average task progress (upstream card reports +17.5% over pi_0.5)Upstream-reported figure on the Hugging Face model card; not independently executed or verified by OSR.

[31]

Project notes

Action space, embodiments, and environment caveats

In config.yml, wall-oss-0.5 configures both continuous flow-matching and autoregressive action horizons (action_horizon_flow: 32, action_horizon_ar: 32, state_bins: 512) with a 26-D default dof_config covering dual-arm relative Cartesian end-effector position, 6D relative rotation, grippers, decomposed base velocity, height, and head actions. Inputs combine multi-view RGB images (448 px in config.yml), proprioceptive state, and language instructions. The model card lists more than 20 supported robot embodiments mapped via norm_key settings—spanning Franka, UR5, ALOHA, and ARX5 as well as benchmarks such as DROID, BridgeData v2, RH20T, and AgiBot World Alpha—so selecting the matching norm_key and state/action dimensions is essential. [1] [30] [31]

Two installation details require reconciliation before setup: the pinned wall-x README documents a Python 3.10 environment, whereas the wall-oss-0.5 model card documents Python 3.12. The README advertises an Ubuntu 22.04 and CUDA 12.x environment and requires a CUDA-capable NVIDIA GPU (a documented upstream environment rather than an independently verified compatibility range), alongside flash-attn 2.8.3, DMuon (for default training configs), a separate --no-deps LeRobot install, and --no-build-isolation editable installation to compile included CUDA operators against the active PyTorch/CUDA stack. Minimum inference VRAM and control frequency are not stated as single scalars in the upstream docs. [1] [30] [31]

As documented upstream starting points (not executed in this review), basic action inference is invoked with python scripts/fake_inference.py --checkpoint-path <path/to/checkpoint>. For fine-tuning on LeRobot-formatted datasets, the README directs users to workspace/README.md and the template at workspace/example/lerobot/qwen2_5_lerobot_template.yml before running python -m wall_x.trainer.fsdp_trainer.train_fsdp --config <path/to/config.yml>. [1]

Upstream-reported benchmarks and unverified runtime

The wall-oss-0.5 model card reports non-trivial zero-shot completion on a 17-task real-robot suite (including held-out deformable manipulation) and 60.5% average task progress after fine-tuning on 15 real-robot tasks (+17.5% over π0.5). All benchmark results are strictly upstream-reported; Open Source Robotics has not compiled CUDA operators, downloaded or executed checkpoints, run training or inference, or evaluated physical robot behavior. [1] [31]

What we checked

Source research only. Selected implementation/model text, repository tree and license declarations were inspected. No upstream compilation, dependency installation, simulator, network/USB/device integration, data collection, benchmark or physical robot test was performed. Source files are linked rather than redistributed; external services and paid products are not endorsed.

Not yet established

  • Documented closed-loop control frequency (Hz)

Updates

Upstream last updated
Sep 18, 2026 [32]
Latest release
Not established
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
GitHub x-square-robot/wall-x@ce2a1ae31d9ee1028b97ca40e3f66034cf8c0e5f; HF x-square-robot/wall-oss-0.5@b7dc23a82dbf70a65859c9fdd596703234597028 · combined hardware and software repository

Separate license matrix

Distinct robot-facing model contribution with actual implementation/model content and a bounded documented next step. Review covers source artifacts, not operational performance or a complete tested deployment.

LayerAvailability / license
Codeavailable · Apache-2.0 code; checkpoint rights separate
Declaration for the scoped source implementation/documentation; dependencies, external services, model/data assets and inherited file-specific terms require separate checks. No blanket redistribution or operational clearance. [1] [2] [3]
Model weightsavailable · License not fully assessed
Ungated 8.33 GB safetensors checkpoint (x-square-robot/wall-oss-0.5) verified at HF revision b7dc23a82dbf70a65859c9fdd596703234597028; cardData has no license declaration, so weight terms remain unverified and separate from repository Apache-2.0. [30] [31]
Training datasetsunavailable · License not fully assessed
Pretraining combines 20+ open-source robotics benchmarks (individual upstream licenses) with proprietary/internal robot data (ex_normal, x2_normal) not redistributed as a unified open dataset. [31]
Documentationavailable · Apache-2.0 code; checkpoint rights separate
Declaration for the scoped source implementation/documentation; dependencies, external services, model/data assets and inherited file-specific terms require separate checks. No blanket redistribution or operational clearance. [1] [2] [3]

Separate code, checkpoint, and dataset rights

The x-square-robot/wall-x repository carries an Apache-2.0 license (GitHub last pushed 2026-09-18). However, the x-square-robot/wall-oss-0.5 checkpoint card and Hugging Face API metadata declare no weight license (cardData: None), and the pretraining corpus mixes 20+ open-source benchmarks with proprietary/internal robot configurations (ex_normal, x2_normal). Code Apache-2.0 does not automatically settle weight or training-data reuse rights. [2] [30] [31] [32]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history

2026-11-21 — Migrated from 0.7.0-draft.5 project to 0.8.0-draft.1 model; added verified HF wall-oss-0.5 parameter count (4.16B), 8.33 GB safetensors checkpoint, action space, and upstream-reported benchmarks.