---
title: InternVLA-A1.5 VLA Action Explorer
emoji: 🤖
colorFrom: red
colorTo: green
sdk: gradio
app_file: app.py
pinned: false
python_version: "3.10"
short_description: Predict robot action chunks from an image + instruction
startup_duration_timeout: 45m
---

# 🤖 InternVLA-A1.5 · VLA Action Explorer

Interactive demo for [**InternRobotics/InternVLA-A1.5-base**](https://huggingface.co/InternRobotics/InternVLA-A1.5-base),
a Vision-Language-Action (VLA) robot policy from the paper
*"InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization."*

The model couples a **Qwen3.5-2B** vision-language backbone with a lightweight
**flow-matching action expert**. Given a single camera frame and a natural-language
task instruction, it predicts the next **50-step continuous action chunk** — the
robot's planned motion — which this Space visualizes dimension by dimension.

## Notes

- Uses the `google_robot` single-arm embodiment (single RGB camera, 7-DoF arm + gripper).
- The video-foresight (WAN2.2) branch is discarded at inference (`action_loss_only=True`),
  so no video-generation weights are downloaded.
- Predicted actions are shown in the model's raw (normalized) flow-matching output space.
- The custom Qwen3.5 modeling code is patched into `transformers` at startup and runs
  through its pure-PyTorch fallback (no flash-attn / flash-linear-attention required).