SAGES

Codebase for SAGES (Semantic Autonomous Guidance Engine for Space). In this repo, the codebase around the free-flyer scenario is released.

Paper: Takubo et al., "Language-Conditioned Safe Trajectory Generation for Spacecraft Rendezvous," AIAA Journal of Guidance, Control, and Dynamics (Under Review).

Webpage: Link

Interactive Demo Video: Link

Code Organization

freeflyer/
├── dataset/                  # Processed datasets (PyTorch format)
│   └── torch/
├── dataset_generation/       # Dataset generation scripts and results
│   └── UMAP_Results/
├── decision_transformer/     # Causal transformer model and training
│   └── saved_files/
├── dynamics/                 # Spacecraft dynamics models
├── optimization/             # SCP-based trajectory optimization
│   └── saved_files/
│       └── warmstarting/
└── test/                     # Test scripts

How to Use

SAGES has three main components in the workflow: (i) dataset generation, (ii) training, and (iii) test-time inference / analysis. This repo is built using poetry, but feel free to switch to another package manager (e.g., uv).

1. Text-Trajectory Dataset Generation

poetry run python freeflyer/dataset/dataset_pargen.py

This generates the dataset of text-trajectory pairs for training and validation by solving a batch of trajectory optimization problems via SCP.

2. Training

poetry run python freeflyer/decision_transformer/main_train_lang.py

This initiates training of the causal transformer, which is a core component of SAGES.

3. Analysis

poetry run python freeflyer/optimization/warmstarting_analsyis.py
# For Jetson deployment:
# poetry run python freeflyer/optimization/warmstarting_analsyis_orin.py

This conducts a warm-start analysis of SAGES, comparing the performance of SAGES (transformer-based warm-starting + SCP refinement) against convex-based warm-starting, given the same terminal state from the warm-start trajectory.