This dataset contains 1,000 verified trajectories generated by the Spectrum Physics Engine, including 158 "Golden Path" soft landings with touchdown velocities as low as 0.01 m/s.

Dataset Description

Overview

This dataset provides physically accurate Mars EDL (Entry, Descent, Landing) trajectory data for training autonomous landing systems, reinforcement learning agents, and physics-based world models.

Key Statistics:

Total Trajectories: 1,000 episodes
Total Frames: 58,620 timesteps
Golden Paths: 158 (touchdown velocity < 3.0 m/s)
Best Landing: 0.01 m/s (near-perfect touchdown)
Source Inventory: 178,857 simulated trajectories
Success Rate in Source: 0.31% (556/178,857)

Physics Environment

Gravity: 0.38g (Mars surface)
Atmosphere: CO2-dominant, variable density (MarsGRAM model)
Simulation Engine: Genesis Physics (GPU-accelerated, Apple Metal)
Time Resolution: 0.1s per frame

Use Cases

Imitation Learning: Clone optimal landing trajectories for autonomous descent
Inverse RL: Learn reward functions from Golden Path demonstrations
World Model Training: Pre-train physics prediction models (Cosmos, Sora-style)
Anomaly Detection: Train discriminators to identify failure modes
Sim-to-Real Transfer: Validate against real Mars mission data

Dataset Structure

mars-edl-golden-path/
├── data/
│   └── chunk-000/
│       └── file-000.parquet
├── meta/
│   └── info.json
├── README.md
└── LICENSE

Parquet Schema

Column	Type	Description
`trajectory_id`	string	Unique trajectory identifier
`trajectory_hash`	string	SHA-256 proof of trajectory
`timestamp`	float64	Time in seconds
`observation_0-11`	float32	State vector (see below)
`action_0-3`	float32	Control inputs
`reward`	float32	Normalized reward (0-1)
`is_terminal`	bool	Episode end flag
`is_success`	bool	Successful landing flag
`episode_index`	int32	Episode number
`entropy_grade`	float32	Normalized landing quality
`final_velocity`	float32	Touchdown velocity (m/s)

Observation Vector (12 dims)

Index	Name	Description
0-2	pos_x, pos_y, pos_z	Position (meters)
3-5	vel_x, vel_y, vel_z	Velocity components (m/s)
6	altitude	Height above surface (m)
7	velocity	Total velocity magnitude
8	rho	Atmospheric density
9	phase	0=DESCENT, 1=POWERED_DESCENT
10-11	padding	Reserved

Action Vector (4 dims)

Index	Name	Description
0-2	thrust_x/y/z	Thrust vector (normalized)
3	throttle	Engine throttle (0-1)

Compatibility

LeRobot: Compatible with lerobot.common.datasets v3.0 schema
Isaac Lab: Parquet format works with Isaac Lab data pipelines
NVIDIA Cosmos: Suitable for physics world model fine-tuning
PyTorch/JAX: Standard Parquet, loadable with pandas/pyarrow

Loading Example

import pandas as pd
from datasets import load_dataset

# Via HuggingFace datasets
dataset = load_dataset("spectrum-ai/mars-edl-golden-path")

# Via pandas
df = pd.read_parquet("data/chunk-000/file-000.parquet")

# Filter Golden Paths only
golden_paths = df[df["is_success"] == True]

Provenance

Generator: Spectrum Physics Engine (DeepGenesis Core)
Hardware: Apple Silicon (M-Series, Metal acceleration)
Date Generated: December 2025 - January 2026
Verification: SHA-256 trajectory hashing
Source: Protocol Company physics inventory

Commercial Licensing

This dataset is available under a commercial license.

Sample/Research Use: Contact for academic access Commercial Use: Contact for licensing terms

Contact: Protocol Company

Citation

@dataset{spectrum_mars_edl_2026,
  title = {Mars EDL Golden Path Trajectories},
  author = {Protocol Company},
  year = {2026},
  publisher = {Hugging Face},
  note = {High-fidelity Mars Entry, Descent, Landing simulation data}
}

LeRobot - Robotics library
NVIDIA Isaac Lab - Robot learning framework
NVIDIA Cosmos - World foundation models

"Physics is not a suggestion. It is a geometry."

Generated by the Spectrum Sovereign Engine | January 2026

Downloads last month: 13

You need to agree to share your contact information to access this dataset

Mars EDL Golden Path Trajectories