Where Does Reinforcement Learning Help Lunar Terrain-Relative Landing? Residual Control vs. Decision-Level Policies
Abstract
Terrain-relative navigation (TRN) lets a lander correct its position estimate against an onboard map, and a classical guidance
baseline with a TRN filter already lands precisely in nearly all default cases, leaving little room for learning. We study where
a learned policy could add anything, using an open 3-DoF lunar landing benchmark built from Lunar Orbiter Laser Altimeter
(LOLA) polar digital elevation models (Site 04, Site 23, Haworth) with a geographic held-out split, a full navigation filter in the
loop, and a realism layer. The remaining headroom appears only at large initial position error, where failures come from missing
terrain fixes rather than from flying, and classical changes to the filter and sensor recover most of the loss: a terrain-match search
window that covers the initial uncertainty, and a camera field of view of at least 40◦ for the remainder. We then compare two uses
of reinforcement learning against these fixes. Residual proximal policy optimisation (PPO), which corrects the thrust command of
the baseline, did not improve it. A decision-level policy, which chooses whether to continue, hold, or detour, has real headroom:
a one-line detour rule raises precise landings by 11 to 12 points on a synthetic blackout benchmark, and a learned offline chooser
matches the rule but does not exceed it, while sampled PPO policies fall below it. We release the benchmark, code, and negative
results. Submitted to Acta Astronautica, October 2026. Code and benchmark: https://github.com/Weirdnemo/lunar-trnrl
Related
- Continuous-Time Reinforcement Learning for N-Player Stochastic Differential Games with Exploratory Policies
- Reinforcement learning on the EPRI AI-ccelerating Unit Commitment benchmark: diagnosing over-commitment and guiding MILP with learned policies
- NeuroSymbolic-RLNet: a neuro-symbolic reinforcement learning framework for transparent sequential decision-making
- Assessing reinforcement learning and vigor with an online decision-making battery
- Impaired Reinforcement Learning Underlying Explore-Exploit Decision Making in Theft Recidivists
BibTeX
@misc{chauhan2026where,
title = {{Where Does Reinforcement Learning Help Lunar Terrain-Relative Landing? Residual Control vs. Decision-Level Policies}},
author = {Nimesh Chauhan},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23642853},
url = {https://zenodo.org/records/23642853},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Where Does Reinforcement Learning Help Lunar Terrain-Relative Landing? Residual Control vs. Decision-Level Policies
AU - Nimesh Chauhan
PY - 2026
DA - 2026/10/25
PB - Zenodo
DO - 10.5281/zenodo.23642853
UR - https://zenodo.org/records/23642853
AB - Terrain-relative navigation (TRN) lets a lander correct its position estimate against an onboard map, and a classical guidance baseline with a TRN filter already lands precisely in nearly all default cases, leaving little room for learning. We study where a learned policy could add anything, using an open 3-DoF lunar landing benchmark built from Lunar Orbiter Laser Altimeter (LOLA) polar digital elevation models (Site 04, Site 23, Haworth) with a geographic held-out split, a full navigation filter in the loop, and a realism layer. The remaining headroom appears only at large initial position error, where failures come from missing terrain fixes rather than from flying, and classical changes to the filter and sensor recover most of the loss: a terrain-match search window that covers the initial uncertainty, and a camera field of view of at least 40◦ for the remainder. We then compare two uses of reinforcement learning against these fixes. Residual proximal policy optimisation (PPO), which corrects the thrust command of the baseline, did not improve it. A decision-level policy, which chooses whether to continue, hold, or detour, has real headroom: a one-line detour rule raises precise landings by 11 to 12 points on a synthetic blackout benchmark, and a learned offline chooser matches the rule but does not exceed it, while sampled PPO policies fall below it. We release the benchmark, code, and negative results. Submitted to Acta Astronautica, October 2026. Code and benchmark: https://github.com/Weirdnemo/lunar-trnrl
ER -