Adaptive Sampling for Active Perception and Mapping

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 164 papers66 conceptsfrozen 2026-10-04

Traditional sensing is largely a passive affair. Whether it is a LiDAR sensor spinning at a fixed rate on an autonomous vehicle or a microscope scanning a biological sample pixel-by-pixel, most sensors follow a "scene-agnostic" schedule. They allocate the same amount of time, energy, and resolution to a blank wall as they do to a complex obstacle.

This uniform approach is inherently inefficient. In most real-world environments, information is not distributed evenly; it is concentrated in "hotspots," such as object boundaries, high-variance gradients, or dynamic regions. Adaptive sampling (or active perception) replaces fixed scanning patterns with a real-time feedback loop: the system uses its current knowledge to decide where to measure next, prioritizing regions that maximize information gain.

The Inefficiency of Uniformity

The fundamental problem with standard sampling is the trade-off between resolution and resource consumption. To capture fine details in a complex area using a uniform grid, one must increase the sampling density across the entire field of view. As Mohammad H. Rahimi and Gaurav S. Sukhatme note in Adaptive Sampling for Environmental Robotics, this wastes significant energy on redundant data in low-variance regions.

In specialized fields like electron microscopy or X-ray spectroscopy, this waste is even more costly. Yaron Meirovitch and Core Francisco Park demonstrate in SmartEM that biological samples are spatially heterogeneous; using a fixed dwell time for every pixel results in massive time loss. Similarly, Kanta Ono argues that uniform energy scans in spectroscopy spend too much time on low-information background areas, potentially damaging samples with unnecessary radiation.

Guiding the Sensor: Priors and Uncertainty

How does a system "know" where to look before it has actually measured the data? Research in this area generally follows two strategies: prior-based guidance and uncertainty-driven exploration.

1. Leveraging Priors

Many systems use a secondary, "cheaper" sensor to guide a more expensive one. For instance, Qiqin Dai and Oliver Cossairt use a standard RGB image to predict an "Importance Map" for a LiDAR sensor. In Adaptive Illumination based Depth Sensing using Deep Learning, they co-optimize a sampling network and a reconstruction network, allowing the system to capture high-fidelity depth maps using as little as 0.06% of the total pixels.

Other priors include temporal continuity—using the previous frame to predict the next, as seen in Amit Shomer’s SampleDepth—or global map data. Jianping Li shows that OpenStreetMap data can be used to slow down a motorized LiDAR to "stare" at informative building facades, improving localization in Adaptive Motorized LiDAR Scanning Control.

2. Quantifying Uncertainty

When priors are unavailable, systems use statistical models to identify where they are most "confused." Eyal Gofer and Guy Gilboa propose using the variance of an ensemble of depth predictors to guide LiDAR pulses in Adaptive LiDAR Sampling and Depth Completion Using Ensemble Variance. By sampling where predictors disagree most, they achieve high accuracy with 4 to 10 times fewer measurements.

In environmental monitoring, Gaussian Processes (GPs) are the standard tool for this. Andrea Jeradi and Alessandro Farinelli use GPs to help robotic watercraft prioritize "hotspots" like pollution plumes, moving away from static "lawnmower" search patterns toward paths that maximize information gain.

Hardware for Active Perception

The shift toward adaptive sampling has necessitated new hardware paradigms. Traditional rotating LiDARs are being challenged by steerable sensors. Karel Zimmermann and Tomás Svoboda explore this in Learning for Active 3D Mapping, where they use solid-state lidars (SSLs) that allow individual depth-measuring rays to be steered in real-time.

Another radical approach is the "light curtain" proposed by Siddharth Ancha and Srinivasa G. Narasimhan. Instead of scanning the whole volume, the sensor only detects objects that cross a specific, user-defined 3D surface. This "bionic" approach, also echoed in the work of Ruixuan Chen on Integrated adaptive coherent LiDAR, mimics the foveated vision of animals, concentrating high-resolution data only on critical regions of interest.

Why It Works: The Information Bottleneck

The success of adaptive sampling across diverse fields—from 6G network mapping to brain connectomics—stems from its ability to bypass the "information bottleneck." By treating sensing as an optimization problem rather than a data-collection task, these methods ensure that every laser pulse, X-ray photon, or robot movement contributes the maximum possible utility to the final model.

As Cyrill Stachniss and Wolfram Burgard demonstrated early on with Coverage Maps, moving beyond binary representations to probabilistic, information-theoretic strategies allows robots to build more accurate maps with fewer measurements, effectively doing "more with less."

Papers in this version

Showing 60 of 64 papers.

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_adaptive_sampling_for_active_perception__bed76e4f,
  title = {Adaptive Sampling for Active Perception and Mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-sampling-for-active-perception-and-mapping/rdr_bed76e4f31754e528dea76f01841ac3f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_adaptive_sampling_for_active_perception__bed76e4f_v1,
  title = {Adaptive Sampling for Active Perception and Mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-sampling-for-active-perception-and-mapping/rdr_bed76e4f31754e528dea76f01841ac3f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}