Does the memory actually enable long-horizon planning and adapt to changing goals?
A fair evaluation must test whether the memory helps the robot plan beyond its immediate view, not just react to what it sees. RAVEN's persistent 3D voxel-ray map allowed aerial robots to perform long-horizon searches in unstructured outdoor environments, and it beat reactive baselines by 85.25% in simulation [1]. That improvement shows that a memory that stores spatial-semantic information over time is what enables the robot to avoid short-sighted decisions. A fair test would include tasks that require the robot to revisit or reason about places it saw earlier, not just navigate to a visible goal.
Adaptability is equally important: the memory should support new goals without rebuilding the map. LagMemo builds a language-3D Gaussian Splatting memory during one exploration and then efficiently queries it for new multi-modal, open-vocabulary goals, significantly outperforming state-of-the-art methods in multi-goal navigation [2]. CityNavAgent similarly uses a global memory of historical trajectories to simplify navigation to previously visited targets, which is a key feature for long-horizon tasks [5]. A fair evaluation should include sequential task changes or new goal queries after the initial mapping, to see if the memory can be reused effectively.
How much memory and computation does it cost, and does it scale to large environments?
A persistent 3D semantic memory is only useful if it doesn't blow up the robot's resources. SceneVGGT keeps GPU memory under 17 GB regardless of input sequence length, using a sliding-window pipeline that aligns local submaps [3]. This is a critical metric: a fair evaluation should measure peak memory usage and processing speed, not just accuracy, because a memory that is too heavy to run in real time won't be deployable. The same paper also shows that storing timestamps and instance-level identities enables change detection, which is a valuable capability for real-world navigation [4].
Scaling to large environments is another key dimension. RAVEN combines short-range voxel search and long-range ray search to handle large outdoor spaces, and it was tested in 10 photorealistic outdoor environments over 100 semantic tasks [1]. A fair evaluation should include environments of varying size and complexity, and measure whether the memory degrades gracefully as the map grows. CityNavAgent also addresses the exponentially expanding action space in long-horizon aerial navigation by decomposing tasks into sub-goals, which is a strategy that a memory-based system should support [5].
Does it work in the real world, including messy and changing environments?
Simulation success is not enough; a fair evaluation must include real-world tests. RAVEN demonstrated real-world applicability through deployment on an aerial robot in outdoor field tests [1]. Similarly, SceneVGGT was evaluated on custom datasets designed for assistive navigation scenarios, showing applicability to real-world situations [4]. A fair evaluation should include field tests or at least photorealistic simulations that capture sensor noise, dynamic obstacles, and lighting changes.
Terrain awareness is a specific real-world challenge that persistent semantic memory can address. In agricultural settings, a system that classifies terrains into six drivability levels and integrates them into the costmap reduced the false positive rate of forbidden regions by 37.79% compared to LiDAR-only navigation [6]. This shows that semantic memory isn't just about finding objects; it's also about safe path planning. A fair evaluation should measure not only task success but also safety metrics like collision rates and forbidden-region violations, especially in environments with diverse terrain types.
About These Sources
This answer is built on 6 peer-reviewed studies — published from 2025 to 2026, 6 from 2024 or later, collectively cited 58 times — selected as the most relevant from 6 studies that passed quality screening, drawn from 49 papers retrieved from a database of over 500 million.
Sources used in this answer
RAVEN: Resilient Aerial Navigation via Open-Set Semantic Memory and Behavior Adaptation
RAVEN's persistent 3D voxel-ray map enabled long-horizon aerial navigation, outperforming baselines by 85.25% in simulation across 100 tasks in 10 photorealistic outdoor environments, and was validated in real-world field tests.
LagMemo: Language 3D Gaussian Splatting Memory for Multi-modal Open-vocabulary Multi-goal Visual Navigation
LagMemo's language 3D Gaussian Splatting memory supported multi-modal open-vocabulary and multi-goal navigation, significantly outperforming state-of-the-art methods on the GOAT-Core benchmark.
SceneVGGT: VGGT-based online 3D semantic SLAM for indoor scene understanding and navigation
SceneVGGT's sliding-window SLAM pipeline kept GPU memory under 17 GB regardless of input length, enabling memory-efficient semantic mapping with competitive point-cloud performance on ScanNet++.
Building temporally coherent 3D maps with VGGT for memory-efficient Semantic SLAM
The VGGT-based framework stored timestamps and instance-level identities for temporal consistency and change detection, and was validated on custom assistive navigation datasets.
CityNavAgent: Aerial Vision-and-Language Navigation with Hierarchical Semantic Planning and Global Memory
CityNavAgent's hierarchical semantic planning and global memory of historical trajectories achieved state-of-the-art performance in aerial vision-and-language navigation in continuous city environments.
Semantic-Aware Navigation for Agricultural Robots via Multi-Category Terrain Classification and Persistent Memory
A semantic-aware navigation system for agricultural robots classified terrains into six drivability levels and reduced false positive forbidden-region rates by 37.79% compared to LiDAR-only navigation.