search

ARCLab-MIT / kspdg

ARCLab-MIT participation in the KSPDG challenge
MIT License
4 stars 0 forks source link

readme

MIT-ARCLab participation in KSPDG challenge

See the information about the challenge in the upstream repository

Resources

See papers and software tools related to this project in the resources folder

Contact

Feel free to reach out if you have questions, feedback, or inquiries!

Environments / Challenge Scenarios

KSPDG is collection of orbital challenge problems defined within different scenario groups (i.e. "evaluation environments" to use nomenclature from reinforcement learning literature). As of Sep 2023, the challenge problems thus far implemented in KSPDG can be grouped into the following types of scenarios

Agents Overview

MIT-ArcLab has developed a diverse range of agents during this research, categorized mainly into two types: Few-Shot Prompting Agents and Fine-Tuning Agents. Each category represent a different strategy in addressing this challenge.

You can explore and learn more about these agents in our agents folder.

Feel free to delve into the folder for detailed information on each agent, including their design, capabilities, and examples.

Data overview

The data associated with our research is organized into three main sections, each serving a unique purpose in the development and evaluation of our agents. Find all the scripts and folders here.

Weights & Biases (wandb) Integration

Human Gameplay Data

Evaluation Results

Explore each section for detailed information, including methodologies, data formats, and access instructions.

Citation

Direct link to arXiv preprint and BibTex citation.

arXiv

@article{rodriguez2024language,
title={Language Models are Spacecraft Operators},
author={Victor Rodriguez-Fernandez, Alejandro Carrasco, Jason Cheng, Eli Scharf, Peng Mun Siew, Richard Linares},
journal={arXiv preprint arXiv:2404.00413},
year={2024}
}

Installation

The installation process includes several components:

:warning: Note These instructions have been written and verified on a macOS. They have been partially tested on Ubuntu 18.04 as well. Other operating systems should be similar with deviations on file and directory names

Install KSP & Making History Expansion

  1. Purchase and Download Kerbal Space Program and Making History expansion: https://store.privatedivision.com/game/buy-kerbal-space-program-ksp
    • Make sure to purchase Direct Download / DRM Free Private Division as the platform. Make sure you are buying KSP1 and not the recently released KSP2; none of this will work on KSP2!
    • Download the most recent version of KSP "On Final Approach" Portable (.zip). As of this writing and testing, the most recent version was v1.12.5
    • Download the most recent version of Making History expansion pack. As of this writing the most recent version was v1.12.1
  2. Unzip ksp-osx-1.12.5.zip to desired location; for simplicity, all instructions assume the unzipped KSP folder is placed on the Desktop
  3. Attempt to open the KSP game executable/app (e.g. KSP.app on Mac)

:warning: Troubleshooting

  • On a Mac, you might run into an error where you can't open the KSP app because the developer can't be verified.
  • To change these preferences on your Mac, choose Apple menu > System Preferences, click Security & Privacy
  • Under the General tab there should be a notification saying something like "KSP.app was blocked ..." if you've attempted to open the KSP app. Click the "Open Anyway" button next to the notification

:warning: Troubleshooting

  • On a Mac, after enabling KSP to be opened in Security and Privacy, you may encounter a bug where the game loading screen stalls indefinitely
  • The workaround is to move the KSP.app icon onto the desktop and then back into the KSP_osx directory. For some reason bash commands didn't seem to work to fix this bug. Had to manually open Finder, drag the KSP.app icon onto the Desktop, and then drag it back into the KSP_osx/ directory
  1. Unzip KSP-Making_History_Expansion-en-us-mac-1.12.1.zip
  2. Follow the instructions in the Instructions-xxx-xx-xx.txt file located in the unzipped Making History Expansion directory.

Instructions:

  1. Copy the two other files located in this folder (.command and .zip) to the folder where the KSP app is located
  2. Once you have copied the files, double click the .command file Thats it! Enjoy the Making History Expansion of Kerbal Space Program!
  1. Test installation by opening KSP (e.g. KSP.app on Mac). When main screen has loaded, select Start Game and you should see options for Play Missions and Mission Builder to confirm that the Making History Expansion was successfully installed

Install KSPDG Mission Save Files

For each differential game environment there are associated mission files created using the "Making History" expansion pack that serves to populate the KSP game engine with the necessary spacecraft in the appropriate orbits. There is also a number of mission save files for in-game software development testing purposes.

The save files are located in this repo under ksp_files/saves/missions and ksp_files/Missions; both sets are necessary to populate the differential game environments. These mission save files must be downloaded and manaully installed into the correct directory of the KSP game installation.

Copy the contents of ksp_files/saves/missions/ and ksp_files/Missions directory into your local installation of KSP. For example, on a Mac with this repo and KSP's install directories on the desktop this would look like:

git clone git@github.com:mit-ll/spacegym-kspdg.git
cd spacegym-kspdg
cp -r ksp_files/saves/missions/. ~/Desktop/KSP_osx/saves/missions
cp -r ksp_files/Missions/. ~/Desktop/KSP_osx/Missions

Install kRPC Server

kRPC is what allows external scripts and processes (such as python programs) to send commands to and control the KSP game engine

  1. Download latest version of kRPC from the GitHub link on the kRPC Getting Started Page. As of this writing, you should download krpc-0.5.2.zip. NOTE: make sure to download a full version, not just the python package; v0.5.2 is a full version but v0.5.3 is just the python package
  2. Unzip krpc-0.5.2/ folder to ~/Desktop/krpc-0.5.2/
  3. Create a new directory in KSP's GameData directory and move all of the krpc contents there
mkdir ~/Desktop/KSP_osx/GameData/kRPC
mv ~/Desktop/krpc-0.5.2/* ~/Desktop/KSP_osx/GameData/kRPC/

Install PhysicsRangeExtender

By default in KSP, high-fidelity physical simulation of spacecraft is only performed for spacecraft very near to the active spacecraft (e.g. only around 2km). PhysicsRangeExtender allows for better simulation (e.g. thusting maneuvers) of more distant spacecraft.

  1. Clone PhysicsRangeExtender (assumed to be cloned to Desktop in these instructions, but you can put it wherever you like since you will be copying things from the clone to GameData)
  2. Copy necessary subfolder from PhysicsRangeExtender to your KSP install's GameData folder
# clone PhysicsRange Extender locally
cd ~/Desktop
git clone git@github.com:jrodrigv/PhysicsRangeExtender.git

# copy the necessary game data for the mod into your KSP install
mkdir ~/Desktop/KSP_osx/GameData/PhysicsRangeExtender
cp -r ~/Desktop/PhysicsRangeExtender/PhysicsRangeExtender/Distribution/GameData/PhysicsRangeExtender/* ~/Desktop/KSP_osx/GameData/PhysicsRangeExtender/

Install kspdg

If you have not yet done so, clone this repository locally on your machine

git clone git@github.com:mit-ll/spacegym-kspdg.git

To install this package, run:

cd spacegym-kspdg
pip install -e .

For development of this package, we recommend using the conda environment defined in environment.yml. To create and activate this environment, run:

cd spacegym-kspdg
conda env create -f environment.yml
conda activate kspdg

:warning: Troubleshooting

(mach-o file, but is an incompatible architecture (have 'x86_64', need 'arm64e'))
  • The workaround for Mac users with M1 chipsets is described here. For Python 3.9, the workaround entails cloning pyerfa locally, checking out a specific version, and installing in the conda environment
# get pyerfa source code and switch to specific release of pyerfa
git clone --recursive https://github.com/liberfa/pyerfa/
cd pyerfa
git fetch origin
git checkout v2.0.0.1

# install specific version of pyerfa in conda environment
conda activate kspdg
pip install .

Install Luna Multiplayer (LMP)

Future Work

Verify Installation

NOTE: Because the KSPDG environments are inexorably linked to the KSP game engine, many of the library's unit/integration test can only be run when a particular game mission file has been loaded and running. This means that verifying installation and testing during code development is a bit more involved than just a single pytest call

Serverless Tests: Quick test to run without KSP game engine running nor kRPC server connection

cd spacegym-kspdg
conda activate kspdg
pytest tests/serverless_tests/

KSP In-Game Tests: These tests require the KSP game engine to be running, the test-specific mission to be loaded, and a connection to the kRPC server

  1. Start KSP game application.
  2. Select Start Game > Play Missions > Community Created > pe1_i3 > Continue
  3. In kRPC dialog box click Add server. Select Show advanced settings and select Auto-accept new clients. Then select Start Server
  4. In a bash terminal:
cd spacegym-kspdg
conda activate kspdg
pytest tests/ksp_ingame_tests/test_pe1_e1_i3.py

# for additional tests, load a different mission in KSP:
# ESC > Quit to Main Menu > Exit to Main Menu > Play Missions > `lbg1_i2` > Continue
pytest tests/ksp_ingame_tests/test_lbg1_lg0_i2.py

# ESC > Quit to Main Menu > Exit to Main Menu > Play Missions > `sb1_i5` > Continue
pytest tests/ksp_ingame_tests/test_sb1_e1_i5.py
  1. You should see the KSP game reset and focus on a vehicle that then performs several orientation and propulsive maneuvers. The pytest command should then indicate the number of passed tests.

:warning: Troubleshooting If you are using a Mac with an arm64 architecture (e.g. M1 chipset) and recieve an error like (mach-o file, but is an incompatible architecture (have 'x86_64', need 'arm64e')), please refer to instructions in the kspdg library installation section about installing pyerfa from source.

Example: Few-Shot Agent Example

[TO BE DONE]

Example: Fine-Tuned Agent Example

There is a specific folder for the fine-tuned agents that you can find in here.

All these agents depend on GPT models as well as some parameters, these parameters are:

Parameter Description
USE_RELATIVE_COORDINATES This parameter will be used to calculate relative velocity and position.
USE_SHORT_ARGUMENT_NAMES This parameter will be used to save or communicate the observations data using abbreviated names for time, fuel usage, etc.
OPENAI_API_KEY This parameter is mandatory for establishing a connection with OpenAI API.
SCENARIO This will determine the scenario you will be running the agent in.
MODEL Given the user's API key there can be multiple models with their unique ID; this parameter represents this ID.
IGNORE_TIME If you don't want to keep the time in the logs and prompts.
SLIDING_WINDOW_SIZE The size of the sliding window that will be used by the model.
SLIDING_WINDOW_STRIDE The stride of the sliding window (skips) for each call.
SYSTEM_PROMPTS All of the different system prompts for each strategy or scenario.
CHAIN_OF_THOUGHT If you want the model to reason with Chain of Thought, this would be added in the system prompt.

Once all the parameters are set, the agent runs as any other agent via kRPC. Select the desired mission in the game and run the script having the kRPC server open.

conda activate kspdg
cd arclab_mit/agents/fine-tuned_agent
python3 fine_tuning_agent.py

Example: Agent-Environment Evaluator

This example walks through how to evaluate agents for scoring purpose. Due to the GUI-based interface of KSP that requires manual interaction, there is not a straight-forward process for hosting a centralized evaluation server upon which participants can submit agents. Instead there is a decentralized process where agents are evaluated locally on particapants' own computers. Participants will then upload their evaluation results to a centralized scoreboard where they will be authenticated and ranked against other participants.

:warning: Honor System While we have taken some steps to prevent falsification of results, there is still potential for cheating. Therefore this competition works largely on the honor system.

If you think you are doing something inappropriate or unfair, you probably are.

We reserve the right to disqualify teams for unsporting behavior

The agent evaluation process uses a compiled python script located in scripts/evaluate.cpython-39.pyc with input arguments in the ordering: <agent-config-file> <environment-module> <environment-class>.

See `evaluation/configs/example_eval_cfg.yaml`` for an example of the config file

Here is a basic example for running an agent-environment evaluation. As with other examples, you begin by:

  1. Start KSP game application.
  2. Select Start Game > Play Missions > Community Created > pe1_i3 > Continue
  3. In kRPC dialog box click Add server. Select Show advanced settings and select Auto-accept new clients. Then select Start Server
conda activate kspdg # while it is not strictly necessary to use conda environments, it is encouraged for development and debugging purpose
cd evaluation # working directory is important due to relative path in cfg.yaml
python evaluate.cpython-312.pyc configs/example_eval_cfg.yaml   # assuming your conda env has python 3.12
                                                                # ohterwise call evaluate.cpython-39.pyc for python 3.9

This should output to a file in the results/ subdirectory with a name like kspdg_results_20231018_125336.txt. That file has JSON-formatted results that look like

{
    "agent_env_results": {
        "is_episode_done": true,
        "closest_approach": 235.2028250841451,
        "closest_approach_time": 200.80000000002892,
        "closest_approach_speed": 77.87944143686991,
        "closest_approach_pursuer_fuel_usage": 651.595703125,
        "pursuer_fuel_usage": 782.56884765625,
        "evader_fuel_usage": 0.0,
        "weighted_score": 983.3304428262093,
        "expected_deltav_at_final_time": 47.97165399572631
    },
    "user_id": "Team Baselines",
    "user_key": "b1bb536a-fe95-4dea-8564-4c8305ac963a",
    "kspdg_version": "0.0.23",
    "agent_name": "Naive-Ned",
    "scenario_environment": "PE1_E1_I3_V1"
}
1313515906

This results file will then be sent to the authentication and scoreboard server for official ranking in the KSPDG Challenge.

INSTRUCTIONS FOR UPLOADING RESULTS TO COMPETITION SCOREBOARD HAVE BEEN EMAILED TO PARTICIPANTS

Cautionary Notes

Here are some things you should NOT do as they will break/invalidate the proper functioning of the KSPDG environments

References

Throughout the documentation and code comments we refer to aerospace literature such as "Vallado Chp 3" for brevity. However this assumes anyone reading this code knows that "Vallado" is short hand for David Vallado's "Fundamentals of Astrodynamics and Applications", which is an unfair assumption to make. Here we list some of our short-hand references