Back to Gap Map Model components / Robotics & embodied AI

X-VLA

Tsinghua University AIR
open weights / Overall score: 2.8

X-VLA is a 0.9B vision-language-action model from Tsinghua University's Institute for AI Industry Research (AIR) and Shanghai AI Laboratory, built on Florence-2. One transformer serves many robot bodies: each data source gets its own small set of learned prompt embeddings. It is released as a pretrained foundation checkpoint with fine-tunes for WidowX, Google Robot, LIBERO, RoboTwin 2.0 and dual-arm cloth folding, and LeRobot supports it natively.

The code and checkpoints are published under the lead author's accounts, which the project page links. LeRobot's own conversions of the checkpoints are not counted toward this entry.

Openness

3 high confidence
3.0
weights
open (X-VLA-Pt and its fine-tunes on the Hub, ungated)
data
described (the cross-embodiment pretraining mixture is not published, though the repository ships readers for DROID, RoboMind and AgiBot World)
code
open (training script, dataset handlers and LoRA fine-tuning code)
license
Apache-2.0 (code and weights)

X-VLA's weights and training code are Apache-2.0. The repository can read the public robot datasets it trained on, but the assembled pretraining mixture is not published.

Adoption

1 medium confidence
1.0

Hugging Face downloads over the trailing 30 days for the foundation checkpoint and five full fine-tunes, most of them for the foundation model. LeRobot's converted copies are downloaded separately and are not counted.

Capability

4 medium confidence
4.0

X-VLA is pretrained across robot bodies and ships two-armed policies, including one for folding cloth, level with GR00T. It has no separate reasoning model and no humanoid control.

  • https://arxiv.org/abs/2510.10274 recorded 2026-11-12

    Abstract: soft prompts as "embodiment-specific prompts" for cross-embodiment training; evaluated across 6 simulations and 3 real-world robots.

  • https://raw.githubusercontent.com/2toinf/X-VLA/main/README.md recorded 2026-11-12

    README: X-VLA-0.9B "achieves state-of-the-art generalization across six simulation platforms and three real-world robots"; X-VLA-RoboTwin2 "for dual-arm coordinated manipulation"; X-VLA-SoftFold "Specialized in deformable object manipulation (e.g., folding and cloth control)".

Verified 2026-11-12, except some scores, still awaiting re-confirmation