Open Source RoboticsRobots, software & learning

X-VLA

Robotics model

880M-parameter (~0.9B) soft-prompted flow-matching Vision-Language-Action model (ICLR 2026) built on Florence-2-large with embodiment-specific learnable soft prompts, Apache-2.0 code and weights, LoRA fine-tuning, and native LeRobot integration. [1] [2] [4] [5]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Robotics researchers; Cross-embodiment VLA engineers; LeRobot practitioners [4]
Parameters
879,738,545 (~0.88B / '0.9B' F32 parameters per safetensors metadata) [2] [5]
Embodiments
Franka Panda (LIBERO, CALVIN, VLABench), WidowX 250 (BridgeData V2 / SimplerEnv), AgileX (RoboTwin 2.0, SoftFold… [4]
Source rights
Named licenses are documented by part. Review licenses

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

Licenses & related hardware
Ecosystem
LeRobot

Requirements · Research details and dates

On this page

Overview

X-VLA (2toINF/X-VLA-Pt and native LeRobot port lerobot/xvla-base) is an 879,738,545-parameter (~0.88B / 0.9B) flow-matching Vision-Language-Action model (ICLR 2026) built on an encoder-only microsoft/Florence-2-large vision-language backbone paired with a 24-layer SoftPromptedTransformer action denoiser. It supports full fine-tuning and LoRA (PEFT) adapters alongside a FastAPI inference server/client. [1] [2] [3] [4]

Model card summary

Architecture

Florence-2-large encoder-only vision-language backbone paired with a 24-layer SoftPromptedTransformer flow-matching action denoiser (hidden_size: 1024, 16 heads, 32 learnable soft prompt tokens across up to 30 domains) and an Action Hub.

Verified from HF model card and config.json.

[2] [3]

Base model

microsoft/Florence-2-large

[2]

Parameter count

879,738,545 (~0.88B / '0.9B' F32 parameters per safetensors metadata)

Exact identical parameter count verified on both 2toINF/X-VLA-Pt and lerobot/xvla-base.

[2] [5]

Input modalities
rgblanguageproprioception

Primary/multi-view RGB images, English language instruction, and proprioceptive state (use_proprio: true).

[3] [4]

Output modalities
ee-pose-absee-pose-deltajoint-positionsgripper-command

Configurable per-embodiment Action Hub (default action_mode: ee6d in X-VLA-Pt config.json).

[3] [4]

Action space

Flow-matching continuous action chunks (num_actions: 30, default action_mode: ee6d with 6D end-effector rotation representation) managed through embodiment-specific Action Hub masks and soft prompts (num_domains: 30).

[3] [4]

Trained / supported embodiments
  • Franka Panda (LIBERO, CALVIN, VLABench)
  • WidowX 250 (BridgeData V2 / SimplerEnv)
  • AgileX (RoboTwin 2.0, SoftFold dual-arm)
  • AgiBot-G1 (IROS 2025 AgiBot World Challenge)
  • Google Robot (SimplerEnv)

[4]

Training datasets
  • bridgedata-v2
  • Heterogeneous cross-embodiment pretraining mixture (LIBERO, CALVIN, VLABench, RoboTwin 2.0, Soft-Fold, Google Robot, AgiBot World)

[2] [4]

Interfaces
  • FastAPI HTTP server (POST /act)
  • Hugging Face Transformers AutoModel (trust_remote_code=True)
  • LeRobot native policy API

[2] [4] [5]

Checkpoints

CheckpointFormatSizeAccessPinned revision
2toINF/X-VLA-Pt (Foundation pretrained)safetensors3.52 GBopenc1c4a64a7e03
lerobot/xvla-base (LeRobot-native port)safetensors3.52 GBopencdb7964e4fe8

Per-checkpoint repository storage is ~3.52 decimal GB (3.28 GiB; 3,519,073,692 bytes in S005 metadata); exact individual safetensors file size is not separated from repository storage in S005.

[2] [5]

Is this for you?

Use cases
  • Cross-embodiment VLA fine-tuning (full or LoRA PEFT)
  • Client-server HTTP policy deployment on real robots
  • Deformable object manipulation and dual-arm coordination

[4]

Known limitations
  • Authors report an ~1% performance drop after converting original checkpoints to Hugging Face format (under active investigation per README)
  • Direct Transformers loading uses trust_remote_code=True

[2] [4]

Safety and limitations
  • Upstream reports an approximately 1% performance drop across datasets after converting checkpoints to Hugging Face format (metric unspecified; cause under investigation at reviewed revision; parity with lerobot/xvla-base not established).
  • Loading 2toINF checkpoints via AutoModel requires trust_remote_code=True (native lerobot/xvla-base avoids custom Hub code); client inference requires matching domain_id and training-space proprioception normalization.

[4] [5]

Compute & fine-tuning

Inference compute
  • Precision: fp32
  • Edge / consumer capable: Yes

3.28 GB F32 safetensors weights; README notes suitability for GPUs, SLURM clusters, or edge devices; exact min_vram_gb and latency_ms not specified.

[2] [4] · Editorial assessment

Fine-tuning

[4]

Host computer

3.28 GB F32 checkpoint designed for GPU servers, SLURM clusters, or edge devices via decoupled client-server inference.

[2] [4]

Operating system
  • Linux

Inferred from documented Linux/CUDA Python environment and shell instructions.

[4] · Editorial assessment

Dependencies
  • Python 3.10
  • PyTorch
  • Transformers
  • FastAPI / json_numpy
  • LeRobot (optional native integration)

[4]

Skills
  • Python
  • PyTorch
  • Hugging Face Transformers / PEFT LoRA
  • FastAPI client-server robotics setup

Inferred from documented Python/PyTorch/CLI setup and training dependencies.

[4] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [4] · Model card / documentation [1] [2]

  • Clone 2toinf/X-VLA and create Python 3.10 conda environment
  • Load 2toINF/X-VLA-Pt (or embodiment checkpoint) via AutoModel.from_pretrained(..., trust_remote_code=True) or use native LeRobot integration (lerobot/xvla-base)
  • Start FastAPI server (model.run(processor, host='0.0.0.0', port=8000)) and query POST /act from robot client

[2] [4]

Upstream-reported benchmarks

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

BenchmarkUpstream-reported scoreScope & caveats
LIBERO / CALVIN ABC_D / SimplerEnv WidowX / RoboTwin 2.0 / IROS 2025 AgiBot World ChallengeLIBERO 98.1%, CALVIN ABC_D 4.43, SimplerEnv WidowX 95.8%, RoboTwin 2.0 70%, VLABench 51.1, IROS 2025 AgiBot World Challenge 1st PlaceUpstream-reported benchmark scores from 2toinf/X-VLA README (fine-tuned checkpoints); not independently verified by OSR.

[4]

Project notes

Checkpoint selection and soft-prompted domain adaptation

Upstream separates the foundation pretrained checkpoint (2toINF/X-VLA-Pt), the native LeRobot port (lerobot/xvla-base), and embodiment-specific fine-tuned checkpoints; choose a foundation checkpoint when fine-tuning on a new robot or task, and use an embodiment-specific checkpoint only when your setup matches that domain. At pinned Hub revision c1c4a64a7e03, config.json specifies hidden_size: 1024, depth: 24, num_heads: 16, num_actions: 30, action_mode: ee6d (6D end-effector rotation), use_proprio: true, num_domains: 30, and len_soft_prompts: 32. Each robot domain uses learnable soft prompts and Action Hub masks; the documented inference client supplies a domain_id, and proprioceptive inputs must match training-space normalization—selecting a domain_id alone does not adapt the model to an unseen robot. [2] [3] [4] [5]

Upstream reports an approximately 1% performance drop across datasets after converting checkpoints to Hugging Face format (the metric is unspecified, and upstream noted the cause was still under investigation at the reviewed revision). Loading 2toINF/X-VLA-Pt via AutoModel requires trust_remote_code=True, whereas lerobot/xvla-base uses native LeRobot code; the supplied evidence does not establish conversion parity or whether the reported conversion drop applies to lerobot/xvla-base. [4] [5]

Footprint and upstream-reported benchmarks

Reported repository storage is approximately 3.52 decimal GB (3.28 GiB; 3,519,073,692 bytes in S005) per ungated FP32 checkpoint (metadata reports total repository storage rather than isolating exact individual weight-file size). Upstream does not pin a single minimum VRAM or closed-loop latency scalar in the README. The repository reports fine-tuned benchmark scores including LIBERO (98.1%), CALVIN ABC_D (4.43), SimplerEnv WidowX (95.8%), RoboTwin 2.0 (70%), VLABench (51.1), and 1st place in the IROS 2025 AgiBot World Challenge. All scores are upstream-reported; Open Source Robotics has not run training, fine-tuning, inference, or hardware evaluations for X-VLA. [2] [4] [5]

What we checked

None. Prepared from public upstream sources without hands-on hardware execution.

Not yet established

  • Documented closed-loop control frequency (Hz)

Updates

Upstream last updated
Oct 20, 2026 [1]
Latest release
Mar 2, 2026 [5]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
May 21, 2027
Reviewed revision
2toinf/X-VLA@6c4ecfb8be458ae7c0c96094e4b0add4174f4fe2; HF 2toINF/X-VLA-Pt@c1c4a64a7e03ac5b95c468bf1578f3d03651b53b; HF lerobot/xvla-base@cdb7964e4fe842935d671bfab5a5ebe00a96648c · software repository

Separate license matrix

Clean Apache-2.0 0.88B soft-prompted flow-matching VLA with ungated safetensors weights, LoRA support, and native LeRobot integration.

LayerAvailability / license
Codeavailable · Apache-2.0
Apache License Version 2.0 verified in root LICENSE. [1] [6]
Model weightsavailable · Apache-2.0
Both 2toINF/X-VLA-Pt and lerobot/xvla-base declare license: apache-2.0; upstream license terms of base model microsoft/Florence-2-large were not excerpted in this review and require independent verification. [2] [5]
Training datasetsavailable · Mixed upstream (heterogeneous cross-embodiment and benchmark datasets; individual upstream dataset terms not excerpted in this review)
Pretraining and benchmark fine-tuning datasets carry separate upstream licenses. [4]
Documentationavailable · Apache-2.0
[4] [6]

Separate code, weights, and pretraining dataset licenses

Both 2toinf/X-VLA (verified in root LICENSE) and both Hugging Face checkpoints (2toINF/X-VLA-Pt and lerobot/xvla-base) carry Apache-2.0 licenses. In contrast, the upstream license terms of the microsoft/Florence-2-large base backbone and the heterogeneous pretraining and benchmark datasets (such as BridgeData V2, LIBERO, CALVIN, RoboTwin, and AgiBot World) were not excerpted in this review and require independent verification. [1] [2] [4] [5] [6]

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 — Initial 0.8.0-draft.1 model draft prepared from pinned GitHub commit 6c4ecfb8be45 and HF revisions c1c4a64a7e03 / cdb7964e4fe8.