DreamLander: A 6-DOF World Model based Reinforcement Learning Planetary Lander

George Nehma, Patrick Quinn, Madhur Tiwari, Camilo Riano-RiosarXiv 2026 · DOI · arXiv

TL;DR

DreamLander is a novel model-based reinforcement learning (MBRL) framework that achieves pinpoint planetary landing accuracy by learning a "world model" of 6-DOF flight dynamics. Built on the DreamerV3 architecture, it successfully navigates realistic, noisy sensor environments to land within 2 meters of a target with soft-landing velocities under 0.3 m/s.

What Problem Does This Paper Address?

Landing a spacecraft on the Moon or Mars—the Entry, Descent, and Landing (EDL) phase—is notoriously difficult. Engineers face a brutal optimization problem: you must deliver a delicate payload to a specific spot while fighting complex gravity and atmospheric forces, all with a severely limited fuel supply.

Traditional Guidance, Navigation, and Control (GNC) systems often rely on simplified models or computationally expensive convex optimization (like the G-FOLD algorithm). While effective, these methods can be rigid. On the other hand, "model-free" reinforcement learning (RL) has shown promise for autonomy but is often sample-inefficient and struggles to generalize to new environments.

The authors of DreamLander aim to bridge this gap by using Model-Based RL, which allows an agent to learn a internal representation of physics—a "world model"—to plan and dream its way to a safe landing.

The 6-DOF coordinate frames used to represent the complex dynamics of the EDL problem.

Methods

The core of DreamLander is DreamerV3, a state-of-the-art MBRL algorithm. Unlike standard RL that learns purely by trial and error in the real world, DreamerV3 learns a Recurrent State Space Model (RSSM). This model maps high-dimensional observations into a compact "latent space" and learns to predict how that space evolves over time.

Key components of the methodology include:

  • Latent Imagination: The agent trains its control policy (the "Actor") and its value estimator (the "Critic") primarily within its own "imagined" trajectories generated by the world model. This makes it incredibly sample-efficient.
  • High-Fidelity Simulation: The system was trained using NVIDIA IsaacLab, leveraging the PhysX engine for realistic 6-DOF dynamics, including collision detection and realistic sensor noise for IMU and LiDAR data.
  • Reward Shaping: To prevent the agent from simply "falling" too fast, the authors used potential-based reward shaping. They introduced a weighted norm that prioritizes correcting horizontal (X-Y) errors over vertical descent, discouraging the lander from crashing before it has aligned with the landing zone.
The DreamerV3 architecture, featuring the World Model (RSSM + Predictor heads) and the Actor-Critic agent.

Datasets

The paper does not use a static dataset; instead, it utilizes a dynamic simulation environment.

  • Environment: A custom Lunar landing scenario in IsaacLab with realistic physical parameters based on the Apollo missions (e.g., 15,000 kg mass, Lunar gravity of -1.62 m/s²).
  • Training: Conducted across 8 parallel environments on a single NVIDIA RTX 3090 GPU for approximately 17 hours.
  • Evaluation: The model was tested on 200 randomized trajectories with varying initial positions, velocities, and orientations to ensure robustness.

Key Insights and Ideas

  • Emergent Control Strategies: One of the most fascinating results was the emergence of an intuitive "tilt-and-fire" strategy. Instead of relying solely on small Reaction Control System (RCS) thrusters for horizontal movement, DreamLander learned to tilt the entire vehicle and use the powerful main engine to "push" itself toward the landing zone—a strategy remarkably similar to those used by SpaceX for booster recovery.
  • Sample Efficiency through "Dreaming": By training on imagined states, the model learns the underlying physics of 6-DOF motion much faster than model-free counterparts.
  • Dynamic Target Regulation: The model proved capable of regulating to any designated landing zone (LZ) coordinates, not just the origin. This is a critical step toward "End-to-End" GNC where a lander might need to pick a safe spot on the fly based on visual data.
Process diagram showing how the model "imagines" future states (h, z) to determine the best action (a) without needing constant real-world interaction.

Key Results

DreamLander demonstrated performance that exceeds the requirements of historical missions:

  • Success Rate: Achieved a 98% success rate in standard evaluation scenarios.
  • Pinpoint Accuracy: Consistently landed within 2 meters of the designated target.
  • Soft Landing: Maintained a landing velocity norm of less than 0.3 m/s, well within the safety margins for delicate scientific payloads.
  • Robustness: The model maintained this accuracy even when subjected to realistic sensor noise (e.g., 0.08m mean noise for LiDAR).

Limitations and Future Work

While highly successful, the authors noted a few hurdles:

  • Terrain Sensitivity: In some tests with uneven terrain, the agent's learned policy of "hovering low" led to premature collisions. Future work will involve further tuning the reward function to encourage safer altitude margins.
  • Pixel-to-Action: Currently, DreamLander uses state-vector observations (position, velocity, etc.). The authors aim to integrate raw camera imagery (pixels) directly into the world model for a truly end-to-end vision-based landing system.
  • Generalization: Testing the "few-shot" adaptation capabilities of the model on different celestial bodies, like Mars or asteroids, remains a promising next step.

Summary updated

Abstract

Soft powered descent landings are becoming evermore important in the field of space exploration and scientific advancement, because of their ability to deliver mission critical payloads, both highly accurately and safely. The high efficiency restriction on planetary landings due to the severely limited fuel supply results in the need for optimal guidance trajectories and control. Model based reinforcement learning is proving to be an important tool in simplifying the highly complex process of generating optimal guidance trajectories and control. In this work, we present DreamLander, a state of the art model based reinforcement learning algorithm based on DreamerV3, to generate optimal guidance and control for a realistic 6 degree of freedom entry descent landing problem. The model based approach creates a sample efficient and highly generalizable model that can be applied to a variety of scenarios whilst maintaining a high degree of accuracy, landing within 2 meters of the designated landing zone, and with velocity norm less than 0.3 meters per second. DreamLander makes use of state sensor observations, making progress toward End-to-End solutions of the Entry Descent Landing problem.
BibTeX
@misc{nehma2026dreamlander,
  author = {George Nehma and Patrick Quinn and Madhur Tiwari and Camilo Riano-Rios},
  title = {{DreamLander: A 6-DOF World Model based Reinforcement Learning Planetary Lander}},
  year = {2026},
  eprint = {2611.12824},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.12824},
}
Figure