GUI Interaction
Complete interface tasks by choosing actions from the instruction and current observation.
The agent reads a task instruction and the current interface state. The policy proposes candidate actions, and the JEPA world model predicts their outcomes to select the next action. Browser feedback provides the updated state for the following decision. In the MiniWoB click-checkboxes task, the agent selects the requested checkboxes and submits the form.
Code and usage
The implementation and usage guides are available on GitHub:
Setup
Use Python 3.13 on Linux with CUDA. With Conda installed, run from the collection root:
conda create -n rsi-gui python=3.13 -y
conda activate rsi-gui
python -m pip install torch==2.6.0 --index-url https://download.pytorch.org/whl/cu124
python -m pip install -r gui_interaction/requirements.txt
python -m playwright install --with-deps chromium
The browser installation includes Chromium and its system libraries. To use an existing Chromium installation, set RSI_CHROMIUM_PATH to its executable.
Download ReasonFlux-Coder-7B and Qwen2.5-Coder-3B-Instruct to RSI_BASE_MODELS_DIR using the backbone download guide.
Run inference
Set OUTPUT_DIR as described in the overview. This command loads the policy adapter and GUI world model for action selection:
CUDA_VISIBLE_DEVICES=0 python gui_interaction/infer.py \
--controller world \
--output "$OUTPUT_DIR/gui_predictions.jsonl"
The command evaluates 100 task seeds. Add --seeds 270000 to run one episode. Outputs include episode outcomes, world-model forecasts, and interaction traces. Use --controller selected to run the policy controller with the same policy adapter.
Checkpoints
| Component | Resource |
|---|---|
| Policy backbone | ReasonFlux-Coder-7B |
| Policy adapter | final/policy |
| World-model backbone | Qwen2.5-Coder-3B-Instruct |
| GUI world model | final/world |