CodeQuasar/orbital-anomaly-openenv

┌──────────────────────────────────────────────────────────────────────┐
│                     MULTI-AGENT LAYER (Theme 1)                     │
│  MissionCommanderAgent  ← oversight, delegation, rationale logging  │
│    ├─ EPSSpecialistAgent       (battery + solar)                     │
│    ├─ ThermalSpecialistAgent   (thermal + payload)                   │
│    └─ CommsSpecialistAgent     (RF chain + comms)                    │
└──────────────────────────────┬───────────────────────────────────────┘
                               │ action
┌──────────────────────────────▼───────────────────────────────────────┐
│              EXTENDED MISSION MODE — 36 STEPS (Theme 2)             │
│  Phase 0 (steps  1-12): EPS Crisis    — solar misalignment + drain  │
│  Phase 1 (steps 13-24): Thermal Crisis — payload heat spike         │
│  Phase 2 (steps 25-36): Comms Crisis  — RF chain degradation        │
│  [battery_soc + payload_temp CARRY OVER between phases]             │
└──────────────────────────────┬───────────────────────────────────────┘
                               │
┌──────────────────────────────▼───────────────────────────────────────┐
│              SPACECRAFT SIMULATOR v3                               │
│  EPS:     power-balance physics (solar input, bus drain, SOC)       │
│  ADCS:    cosine solar alignment, reaction wheel saturation         │
│  Thermal: 3-zone propagation (payload / avionics / battery)         │
│  Comms:   transponder + antenna pointing chain                      │
│  Orbital: eclipse cycles, GS windows, radiation, science windows   │
│  Faults:  13-fault latent causal graph (hidden, cascading)          │
│  Partial observability: 6 sensor dropout patterns                  │
└──────────────────────────────┬───────────────────────────────────────┘
                               │ metadata.fault_beliefs (every obs)
┌──────────────────────────────▼───────────────────────────────────────┐
│              FAULT BELIEF STATE (Theme 3: World Model)              │
│  Heuristic posterior over 13 faults from observable symptoms.       │
│  Updated every step. Agent reasons about hidden fault state.        │
│  mppt_stuck(78%), radiator_stuck(60%), transponder_hot(52%)...      │
└──────────────────────────────────────────────────────────────────────┘

reward = safe_map(
    0.30 × eps_score          # battery SOC + bus voltage
  + 0.22 × thermal_score      # 3-zone temperatures within limits
  + 0.18 × adcs_score         # attitude error + wheel saturation
  + 0.15 × comms_score        # BER + packet loss
  + 0.15 × survivability      # catastrophe multiplier
) × survivability
+ science_bonus(0.12)         # payload ON during observation_window_active

# Survivability: ×0.4 if SOC<10%, ×0.7 if SOC<20%, ×0.5 if avionics>80°C
# Mapping: reward = 0.001 + raw × 0.998  (strict open interval, grader-safe)

Step 0 — Hard task (eclipse, 7 faults active, none directly observable)

  mppt_stuck              ████████████████░░░░  78%  ← EPS fault inferred from low solar
  radiator_valve_stuck    ████████████░░░░░░░░  62%  ← Thermal fault inferred from high temp
  transponder_overheating ██████████░░░░░░░░░░  53%  ← Comms fault inferred from low signal
  battery_aging           ████████░░░░░░░░░░░░  43%
  heat_pipe_failure       ███████░░░░░░░░░░░░░  38%

Step 1 — after switch_power_bus
  mppt_stuck              ████████████░░░░░░░░  60%  ← belief updated as SOC rises
  ...

Phase 0: EPS (1-20)       Phase 1: Thermal (21-40)   Phase 2: Comms (41-60)    Phase 3: Cascade (61-80)
┌──────────────────┐       ┌──────────────────┐       ┌──────────────────┐
│ Steps  1-12      │──────▶│ Steps 13-24      │──────▶│ Steps 25-36      │
│ SOC:  22-38%     │       │ Temp: 79°C+      │       │ BER:  0.18+      │
│ Solar: degraded  │       │ Radiator fault   │       │ Antenna stall    │
└──────────────────┘       └──────────────────┘       └──────────────────┘
         ↑─────────── battery_soc + payload_temp + wheel_sat CARRY OVER ─────↑

from inference import mission_commander_decide, compute_fault_beliefs

obs = env.reset(task_id="hard").observation

# World model: 13-fault belief state
beliefs = compute_fault_beliefs(obs)
# → {"mppt_stuck": 0.78, "radiator_valve_stuck": 0.60, ...}

# Commander delegates to highest-confidence specialist
action, rationale, recs = mission_commander_decide(obs)
# → ("switch_power_bus",
#    "[EPS_Specialist|97%] CRITICAL: battery at floor — reserve bus",
#    {"EPS_Specialist": (...), "Thermal_Specialist": (...), "Comms_Specialist": (...)})

from client import OrbitalAnomalyOpenenvEnv
from models import OrbitalAnomalyOpenenvAction

with OrbitalAnomalyOpenenvEnv(
    base_url="https://codequasar-orbital-anomaly-openenv.hf.space"
).sync() as env:
    result = env.reset(task_id="hard")
    obs    = result.observation

    print(f"Phase: {obs.metadata['phase']} | SOC: {obs.battery_soc:.1f}%")
    print(f"Fault beliefs: {obs.metadata['fault_beliefs']}")

    for step in range(36):
        action = OrbitalAnomalyOpenenvAction(action_type="rotate_to_sun")
        result = env.step(action)
        print(f"Step {step+1}: reward={result.reward:.4f}")
        if result.done:
            break

# Reset
curl -X POST https://codequasar-orbital-anomaly-openenv.hf.space/reset \
  -H "Content-Type: application/json" -d '{"task_id": "hard"}'

# Step
curl -X POST https://codequasar-orbital-anomaly-openenv.hf.space/step \
  -H "Content-Type: application/json" -d '{"action_type": "rotate_to_sun"}'

# State
curl https://codequasar-orbital-anomaly-openenv.hf.space/state

python inference.py                          # multi-agent heuristic
HF_TOKEN=hf_... python inference.py         # LLM via HF inference API

git clone https://github.com/umed-indulkar/orbital-anomaly-openenv.git
cd orbital-anomaly-openenv
uv sync
openenv validate       
uv run server          # → http://localhost:8000

docker build -t orbital-anomaly-openenv .
docker run -p 8000:8000 orbital-anomaly-openenv

python test_reward.py

orbital-anomaly-openenv/
├── server/
│   ├── app.py                                       # FastAPI + OpenEnv server
│   └── orbital_anomaly_openenv_environment.py       # v3 spacecraft simulator
├── models.py                                        # Pydantic typed models 
├── client.py                                        # Typed Python client
├── inference.py                                     # Multi-agent baseline + LLM policy
├── test_reward.py                                   # Complete V2 test suite
├── Orbital_Anomaly_openenv.ipynb                 # GRPO training notebook 
├── blog.md                                          # Plain-English explainer + images
├── images/                                          # All blog/training visualizations
│   ├── task_snapshot.png
│   ├── baseline_distributions.png
│   ├── training_analysis.png
│   ├── action_policy_heatmap.png
│   ├── fault_belief_evolution.png
│   ├── telemetry_timeline_36step.png
│   └── final_dashboard.png
├── openenv.yaml                                     # OpenEnv manifest
├── pyproject.toml                                   # Dependencies (uv)
├── Dockerfile                                       # HF Spaces container
└── README.md

@misc{orbital-anomaly-openenv-2026,
  title   = {Orbital Anomaly OpenEnv: A Spacecraft Digital-Twin Benchmark
             for LLM World Modeling and Long-Horizon Planning},
  author  = {Indulkar, Umed},
  year    = {2026},
  url     = {https://huggingface.co/spaces/codequasar/orbital-anomaly-openenv}
}