Sph3inxz/lewm-pilot-flight

LeWM Pilot — flight fine-tuned world model

State-vector Learned World Model (LeWM) fine-tuned on SkyMind Phase 1 flight telemetry (C172 + T-6 autopilot sessions). Used by the LeWM-Pilot demo planner.

Model summary

Input	52-d normalized flight state
Action	8-d control vector
Environment	5-d weather/context vector
Latent	192-d
Architecture	MLP encoder + 4-layer Transformer predictor
Validation MSE	0.0046 (latent space, holdout)

Files

• lewm_flight_v1.pt — PyTorch checkpoint (model_state_dict, latent_dim, metadata)
• config.json — dimensions and training metadata

Load in SkyMind / LeWM-Pilot

from pathlib import Path
from skymind_core.lewm.engine import LeWMEngine

engine = LeWMEngine(device="cpu")
engine.load_checkpoint("checkpoints/lewm_flight_v1.pt")
latent = engine.encode(obs_vector)          # obs: (52,)
next_latent = engine.predict(latent, action, env)  # action: (8,), env: (5,)

Download from Hub:

hf download Sph3inz/lewm-pilot-flight lewm_flight_v1.pt --local-dir checkpoints

Training

Fine-tuned from a randomly initialized state encoder + predictor on Parquet/Lance flight sessions collected via JSBSim (jsbgym). Config: configs/lewm_finetune.yaml in the GitHub repo.

Demo

Pair with the LeWM-Pilot AI server:

python scripts/run_demo.py

Use --mock-planner if you only want to test the stack without this checkpoint.

Citation / links

• Code: https://github.com/Sph3inz/LeWM-Pilot
• Base LeWM concept: stable-worldmodel / invert-bench lineage