X-VLA is a soft-prompted, flow-matching VLA (arXiv 2510.10274) now integrated into LeRobot. It treats each hardware setup as a task and encodes it with a small set of learnable soft-prompt embeddings, letting one Transformer backbone absorb heterogeneous robots, cameras and action spaces. X-VLA-0.9B is pretrained on 290K episodes across seven platforms, and adapts to new robots by learning new prompts; with about 9M tuned parameters it reports near-π0 performance on LIBERO and Simpler-WidowX. In LeRobot it is used with policy.type=xvla, an action registry (including auto mode) and domain IDs.
| Architecture | Soft-prompted flow-matching VLA built from pure Transformer encoders: learnable per-domain soft-prompt embeddings encode embodiment/camera/domain differences from the earliest fusion stage (24 layers, hidden size 1024, 16 heads, 32-length soft prompts). |
|---|
| Params | 0.9B (X-VLA-0.9B) |
|---|
| Inputs outputs | Inputs: multi-view images, language, proprioceptive state, domain ID; outputs: action chunks (default 32 steps), e.g. 20-D ee6d or 14-D joint actions, auto-padded to the robot's action dimension. |
|---|
| Training data | Phase I pretraining on 290K episodes from DROID, RoboMIND and AgiBot spanning seven platforms and five arm types; Phase II adaptation by learning new soft prompts. |
|---|
| Embodiments supported | Single-arm to bimanual arms; checkpoints for LIBERO, WidowX (Bridge), AgileX, Google Robot, cloth folding, SO-101 bimanual via action modes. |
|---|
| License weights | Apache-2.0 |
|---|
| License code | Apache-2.0 |
|---|
| Hardware requirements | CUDA GPU; bfloat16 recommended for fine-tuning to avoid OOM. |
|---|
| Benchmarks claimed | | Benchmark | Soft-FOLD cloth folding (lerobot/xvla-folding) |
|---|
| Result | 100% success over 2 hours of continuous folding |
|---|
| Source url | huggingface.co/…/xvla |
|---|
|
|---|
| How to try | huggingface.co/…/xvla-base |
|---|
| Sources | |
|---|
Benchmarks, where listed, are as claimed by the vendor.
Illustration
