Document live backend architecture, add sdk_version 6.12.0

README.md

@@ -4,24 +4,41 @@ emoji: 🌍
 colorFrom: gray
 colorTo: red
 sdk: gradio
+sdk_version: 6.12.0
 app_file: app.py
 pinned: false
-short_description: Interactive action-conditioned world rollout from one image
+short_description: Live interactive world rollout from an image
 python_version: "3.10"
 startup_duration_timeout: 40m
 ---
 
 # 🌍 ABot-World — Interactive World Rollout
 
-Turn a single image into a navigable world. Provide a first-frame image, describe
-the scene, and choose an action (WASD to move / IJKL to look). The model
-autoregressively rolls out an action-conditioned video of the world.
+Upload a single starting image and steer a **live** navigable world in real time.
+Provide a first-frame image (image-to-video seed), describe the scene, and drive
+the world with **WASD** (move & turn) / **IJKL** (look & pan). The model
+autoregressively rolls out an action-conditioned world and streams decoded frames
+straight to your browser.
 
 - **Model:** [`acvlab/ABot-World-0-5B-LF`](https://huggingface.co/acvlab/ABot-World-0-5B-LF) (built on Wan2.2-TI2V-5B)
 - **Code:** [amap-cvlab/ABot-World](https://github.com/amap-cvlab/ABot-World)
 - **Project:** [ABot-World](https://amap-cvlab.github.io/ABot-World/)
 
-This Space adapts the upstream real-time interactive web client into a request/response
-rollout suitable for ZeroGPU: instead of live keyboard streaming, you pick one action
-to hold and a rollout length, and receive a video. Inference is causal / block-wise,
-matching the upstream `CausalInferencePipeline`.
+## Architecture
+
+This Space runs a proper live backend/infrastructure (in the spirit of
+[`Overworld/waypoint-1-5`](https://huggingface.co/spaces/Overworld/waypoint-1-5)):
+
+- **`gradio.Server`** exposes ZeroGPU-friendly `/start_game` and `/stop_game`
+  API endpoints plus a raw **WebSocket** (`/ws`) that streams binary JPEG frames
+  and receives live control input — no Gradio UI polling.
+- A custom **`index.html`** front-end handles **image upload** (the i2v seed —
+  no video upload needed), live keyboard steering, and a clean HUD.
+- Every endpoint is keyed by a per-client `session_id`, so concurrent players
+  never share seed images, frame queues, or status messages.
+- **ZeroGPU quota** is billed to the *requesting user*: the HF iframe's
+  `x-api-token` / `x-ip-token` is forwarded by the Gradio JS client and the
+  incoming request context is propagated into the GPU worker thread, so
+  `@spaces.GPU` consumes the visitor's own quota rather than the Space owner's.
+
+Inference is causal / block-wise, matching the upstream `CausalInferencePipeline`.