MAGICIAN: Efficient Long-Term Planning with Imagined Gaussians for Active Mapping

Shiyao Li; Antoine Guédon; Shizhe Chen; Vincent Lepetit

cs.CVMar 23, 2026

MAGICIAN: Efficient Long-Term Planning with Imagined Gaussians for Active Mapping

Authors: Shiyao Li, Antoine Guédon, Shizhe Chen, Vincent Lepetit

Organizations: LIGM, École Nationale des Ponts et Chaussées, IP Paris, Univ Gustave Eiffel, CNRS, France · École Polytechnique, France · Inria, École normale supérieure, CNRS, PSL Research University, France

Abstract

Active mapping aims to determine how an agent should move to efficiently reconstruct unknown environments. Most existing approaches rely on greedy next-best-view prediction, resulting in inefficient exploration and incomplete reconstruction. To address this, we introduce MAGICIAN, a novel long-term planning framework that maximizes accumulated surface coverage gain through Imagined Gaussians, a scene representation based on 3D Gaussian Splatting, derived from a pre-trained occupancy network with strong structural priors. This representation enables efficient coverage gain computation for any novel viewpoint via fast volumetric rendering, allowing its integration into a tree-search algorithm for long-horizon planning. We update Imagined Gaussians and refine the trajectory in a closed loop. Our method achieves state-of-the-art performance across indoor and outdoor benchmarks with varying action spaces, highlighting the advantage of long-term planning in active mapping.

Figures & tables

Figure 2: Overview of the proposed MAGICIAN framework. At time $t$ , we first predict the occupancy field using the occupancy model and update the Imagined Gaussians. We can then efficiently estimate the coverage gain and apply beam search to plan $N_{b}$ candidate trajectories, selecting the one with the highest expected gain. The agent then executes the first $N_{f}$ actions of the best trajectory $\tau_{k}$ of length $N_{d}$ before repeating this process in the next planning loop. In this figure, lighter colors in the Imagined Gaussians indicate higher novelty, while darker colors correspond to previously observed areas. The first trajectory darkens the novelty field the most, representing the optimal path at time $t$ .

Figure 3: Computing coverage gain with Imagined Gaussians. During beam search, we evaluate candidate poses by rendering novelty maps from the Imagined Gaussians to compute the coverage gain. The corresponding depth maps are then used to update the novelty

{\hat{\gamma}}

of Gaussians within a depth tolerance

\epsilon_{d}

Panel of Figure 4 — Figure 4: Evolution of Imagined Gaussians Compared with Ground Truth Mesh. The brighter the Gaussians, the higher their predicted occupancy. As exploration progresses (from left to right), our Imagined Gaussians increasingly align with the ground truth mesh, demonstrating improved environmental modeling.

Panel of Figure 5 — Figure 5: 3D reconstructions obtained with our trajectories. We show Gaussian splatting renderings (top row) and normal maps of the reconstructed meshes (bottom row) after applying Mesh-In-the-Loop Gaussian Splatting [ 17 ] on 100 RGB images collected along our trajectories. The trajectories output by our method cover the entire scene surfaces, resulting in complete and accurate surface meshes.

	MP3D		Macarons++
	Wheeled Robot	Drone	Drone
Action space	Forward (6.5 cm), Turn ( $\pm 10^{\circ}$ )	6 DoF	6 DoF
Camera Height	1.25 m above floor	N/A (Free 3D)	N/A (Free 3D)
Image Resolution	$256\times 256$	$1200\times 680$	$456\times 256$
Field of View	$90^{\circ}\times 90^{\circ}$	$60^{\circ}\times 90^{\circ}$	$60^{\circ}\times 91.6^{\circ}$
Total Steps	$1,000$ or $2,000$	$5,000$	$100$

Table 1: Configuration details on MP3D and Macarons++ benchmarks for active mapping evaluation.

Panel of Figure 6 — Figure 6: Qualitative comparison of novel view synthesis (top row) and surface reconstruction (bottom row) in outdoor and indoor scenes. For each method, we show RGB Gaussian splatting renderings and normal maps of reconstructed meshes after applying Mesh-In-the-Loop Gaussian Splatting [ 17 ] on 100 images collected along the trajectory. The trajectories computed with our method produce more accurate and complete reconstructions, resulting in better rendering quality and preventing holes in reconstructed surfaces.

	AUC $\uparrow$	Final coverage $\uparrow$
Random Walk	0.241	0.324
SCONE [ 19 ]	0.534	0.670
MACARONS [ 20 ]	0.647	0.819
FisherRF [ 26 ]	0.546	0.786
MAGICIAN (Ours)	0.721	0.919

Table 2: Evaluation results on the Macarons++ dataset.

Method	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	Acc. (%) $\uparrow$
FisherRF [ 26 ]	0.55	13.95	0.38	79.15
MACARONS [ 20 ]	0.61	15.68	0.34	86.42
MAGICIAN (Ours)	0.64	17.12	0.30	94.20

Table 3: Evaluation of novel-view rendering and mesh reconstruction on large-scale real-world scanned scenes. Our method achieves the best performance across all metrics.

Setting	Method	Comp. (%) $\uparrow$	Comp. (cm) $\downarrow$
Wheeled Robot	Random	45.67	26.53
	FBE [ 44 ]	71.18	9.78
	UPEN [ 16 ]	69.06	10.60
	OccAnt [ 38 ]	71.72	9.40
	ANM [ 45 ]	73.15	9.11
	NBP [ 32 ]	79.38	6.78

Table 4: Evaluation results on the MP3D dataset. Our method consistently outperforms existing approaches under various robot and action-space settings.

Panel of Figure 7 — Figure 7: Ablation study on the beam search parameters. The horizontal axis denotes the beam width $N_{b}$ , and the vertical axis represents the look-ahead steps $N_{d}$ . Five steps correspond roughly to half the size of the scene.

Figure 8: Ablation study on replanning frequency. The horizontal axis indicates the number

N_{f}

of movement steps executed after each planning phase before replanning. Replanning every 6 steps already provides state-of-the-art results.

Occupancy Model	AUC $\uparrow$	Cov. $\uparrow$
Pretrained	0.652	0.888
Fine-tuned	0.646	0.893

Table 5: Ablation study on comparing models with and without fine-tuning on indoor environments. The fine-tuned version shows a minor 0.5% improvement in final coverage, while the original model retains higher exploration efficiency.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Scene	Rand. Walk	SCONE [ 19 ]	MACARONS [ 20 ]	FisherRF [ 26 ]	MAGICIAN (Ours)
Dunnottar Castle	0.149	0.366	0.618	0.500	0.745
Colosseum	0.219	0.589	0.656	0.551	0.704
Bannerman Castle	0.192	0.559	0.575	0.595	0.761
Pantheon	0.198	0.465	0.601	0.270	0.644
Christ the Redeemer	0.439	0.772	0.859	0.727	0.793
Statue of Liberty	0.323	0.632	0.711	0.553	0.797

Appendix

Table 6: AUCs of full scenes on the Macarons++ dataset.

Scene	Rand. Walk	SCONE [ 19 ]	MACARONS [ 20 ]	FisherRF [ 26 ]	MAGICIAN (Ours)
Dunnottar Castle	0.225	0.527	0.820	0.809	0.975
Colosseum	0.272	0.755	0.794	0.757	0.872
Bannerman Castle	0.240	0.722	0.834	0.801	0.917
Pantheon	0.309	0.610	0.796	0.444	0.842
Christ the Redeemer	0.581	0.924	0.967	0.876	0.973
Statue of Liberty	0.443	0.819	0.909	0.819	0.947

Appendix

Table 7: Final Coverages of full scenes on the Macarons++ dataset.

Figure 9: Standard deviation of the final coverage across different methods and scenes. Our method achieves consistently low values for this metric, indicating strong robustness to random starting poses, whereas other methods exhibit much larger variability.

Panel of Figure 10 — Figure 10: Visualization of exploration trajectories and qualitative comparisons of novel view synthesis and surface reconstruction in outdoor scenes. From top to bottom, the scenes are Pisa Cathedral and Neuschwanstein Castle. In the same scene, all methods start from the same initial camera pose, and for each trajectory visualization, we additionally show the final camera pose at the end of the trajectory. Our trajectory planning method yields more accurate and complete reconstructions, resulting in higher-quality renderings and effectively preventing holes or noise in the reconstructed surfaces.

Panel of Figure 11 — Figure 11: Visualization of exploration trajectories and qualitative comparisons of novel view synthesis and surface reconstruction in indoor scenes. From top to bottom, the scenes are St. Sofia Church and Barts. In the same scene, all methods start from the same initial camera pose, and for each trajectory visualization, we additionally show the final camera pose at the end of the trajectory. Our trajectory planning method yields more accurate and complete reconstructions, resulting in higher-quality renderings and effectively preventing holes or noise in the reconstructed surfaces.

$N_{d}$	10	15	20	25	30	50
AUC	0.652	0.673	0.664	0.662	0.658	0.652
Cov.	0.888	0.887	0.892	0.879	0.881	0.878

Appendix

Table 8: Ablation study on look-ahead steps

N_{d}

Density	$0.5\times$	$1\times$	$2\times$	$4\times$
AUC	0.640	0.652	0.672	0.685
Cov.	0.848	0.888	0.895	0.905

Appendix

Table 9: Ablation study on proxy point sampling density (

1\times

indicates the original density).

Long-Horizon Planning

Long-horizon planning studies how agents coordinate actions over extended sequences to achieve goals amid changing states and uncertainty.

See all 60 papers

TRACE: Ergodic Trajectory Optimization for Active Scene Reconstruction
Aug 3, 2026Ziyue Zheng, Linli Shi, Bingkun He +2Trajectory Optimization Trajectory Generation
1Johns Hopkins University, USA · University of Pennsylvania, USA
All Roads Lead to Rome: Flow-driven Multi-Anchor Exploration for Open-Environment Active 3D Mapping
Sep 29, 2026Yang Li, Aming Wu, Zihao Zhang +33D Reconstruction
School of Artificial Intelligence, Tianjin University, China · School of Computer Science and Information Engineering, Hefei University of Technology, China
Paparazzo: Active Mapping of Moving 3D Objects
Apr 21, 2026Davide Allegro, Shiyao Li, Stefano Ghidoni +1Next-Best-View Planning Informative Path Planning
University of Padova · 2LIGM, ´Ecole Nationale des Ponts et Chauss´ees, IP Paris, Univ Gustave Eiffel, CNRS

Aug 3, 2026J/K move · Enter open · S save

TRACE: Ergodic Trajectory Optimization for Active Scene Reconstruction

Ziyue Zheng, Linli Shi, Bingkun He, Wen Jiang+1

1Johns Hopkins University, USA · University of Pennsylvania, USA

Trajectory Optimization Trajectory Generation 3D Reconstruction Active View Selection Ergodic Control

Open paper AskPDF arXiv

Existing active reconstruction systems with Gaussian-splatting maps select observations greedily, optimizing a single next-best-view (NBV) at each step and connecting the chosen views by short-horizon path planning. This greedy decoupling disregards the global structure of scene information, producing inefficient trajectories that waste sensing capacity in transit between selected views. In this work, we study active reconstruction as an ergodic coverage problem: the time-averaged spatial statistics of the sensor trajectory should match a target information distribution induced by the current map. Our approach derives this target distribution online from uncertainty and visibility, and calculates ergodic trajectories via a kernel-ergodic horizon planner with gradient flow and footprint depletion, closing the loop between mapping and trajectory optimization. We thoroughly evaluate TRACE on the Replica dataset against the Next-Best-View (NBV) baselines, improving PSNR by 1.5 dB. Code: https://github.com/spikelab-jhu/trace-active-reconstruction.

Abstract

Long-Horizon Planning

Explore similar work

TRACE: Ergodic Trajectory Optimization for Active Scene Reconstruction

All Roads Lead to Rome: Flow-driven Multi-Anchor Exploration for Open-Environment Active 3D Mapping

Paparazzo: Active Mapping of Moving 3D Objects