Update run-exp.md

cd VAGEN

bash scripts/install.sh

# go to release branch of verl

cd ../verl

git checkout release

cd ../VAGEN

# Login to wandb for experiment tracking

wandb login

```

## Running Experiments

### Basic Approach

The simplest way to run an experiments is to use the provided scripts:

```

# run one of the experiment scripts

bash vagen/examples/frozen_lake_aico/run.sh  #aico-frozenlake

bash vagen/examples/frozen_lake_trico/run.sh  #trico-frozenlake

bash vagen/examples/frozen_lake_aico/run.sh  

bash vagen/examples/frozen_lake_trico/run.sh  

bash vagen/examples/sokoban_aico/run.sh 

bash vagen/examples/sokoban_trico/run.sh  

```

**Handling Training Instabilities**

Each run takes approximately 4 hours to reach 150 steps on 4 H100s. You can decrease testing frequency to speed up training. Note that training might be unstable due to loss spikes; we recommend restoring from the latest checkpoint when encountering such cases.

### Service Approach

The simplest way to reproduce our experiments is to use the provided scripts:

1. Start the server process in a tmux session 

```

bash vagen/examples/frozen_lake_aico_service/run.sh  # aico-frozenlake

python vagen/server/server.py

```

## Support Environments

The following environments are currently registered:

> NOTICE: SVG and Navigation envs are commented in `env/__init__.py`for better package management. Please comment them out for future use, and refer `vagen/env/README.md` to install dependencies

2. After the server is running, open new terminal to run the training processe:

```

bash vagen/examples/frozen_lake_aico_service/run.sh

bash vagen/examples/svg_aico/run.sh

bash vagen/examples/svg_trico/run.sh

```

## Support Environment

- FrozenLake: A simple grid-based environment

- Sokoban: A visual puzzle-solving environment with box pushing

- SVG: An environment that generate svg code fot provided image. Supports reward model integration