Source-linked AI summary
Leveraging Distributed Sensing and Hyperdimensional Computing for Non-Line-of-Sight Passive Object Sensing
Placido Mursia, Marco Rossanese, Francesco Devoti, Adrian Lendinez Ibanez, Lanfranco Zanzi, Xi Li, Xavier Costa-Perez
TL;DR
Passive entities beyond a robot’s field of view can pose collision risks, while onboard sensing is limited by line of sight and occlusions. The paper uses distributed channel measurements encoded with HDC for risk-aware classification. Ray-tracing results report up to 89% detection accuracy in complex indoor environments, with computational requirements suitable for typical robot hardware.
Problem
Passive entities outside a robot’s field of view can pose safety risks, while onboard LiDAR and camera sensing is constrained by line of sight and occlusions.
Source: §I · p. 1Method
The framework combines distributed infrastructure channel measurements with HDC hypervector encoding to infer passive-object risk without explicit channel reconstruction.
Source: §I · p. 2, §VI · p. 6Results
Up to 89% detection accuracy was demonstrated in complex indoor environments, while maintaining computational requirements on typical robot hardware.
Source: §VI · p. 6Takeaways & Limitations
The system provides preemptive non-line-of-sight information for robot path adaptation as an auxiliary sensing layer that complements existing sensors.
Source: §II · p. 2The proof of concept is binary safe/danger classification; localization, trajectory estimation, and multi-class scene understanding are left for future work.
Source: §II · p. 2
Abstract
from arXiv · showhide
In smart factory environments, passive entities such as humans and movable obstacles may interfere with autonomous robot operations, posing significant safety risks. Detecting potential conflicts between robots and entities outside their field-of-view (FoV) is therefore of paramount importance. However, conventional passive object detection systems rely on sensors such as LiDAR, radar, or cameras, which operate under direct line-of-sight (LoS) and are limited by obstacles and occlusions. To overcome these limitations, integrated sensing and communication (ISAC) enables environmental perception via the communication infrastructure thanks to cooperation among distributed wireless devices. In this paper, we propose a framework for passive entity sensing beyond the robot's FoV that relies solely on distributed channel measurements from the communication infrastructure to perform risk-aware classification. These measurements are embedded into high-dimensional representations using hyperdimensional computing (HDC), enabling robust and scalable learning without explicit channel reconstruction. Numerical simulations based on a commercial ray tracer and ROS-based validation in realistic smart factory scenarios demonstrate improved safety and enhanced path planning with up to 89% accuracy under severe occlusions, while maintaining low computational complexity.
I. INTRODUCTION
In shared industrial spaces, passive entities can create collision risks, while conventional onboard sensing is constrained by line of sight and occlusions. The paper combines distributed wireless sensing with HDC to assess risk beyond a robot’s direct visibility.
I. INTRODUCTION
- Conventional robot-mounted LiDAR and camera sensing is limited by line-of-sight field of view, occlusions, walls, sensing range, and computational capacity.Source: §I · p. 1
- Cooperative distributed channel measurements target regions outside robot visibility, including passive objects around corners or behind shelving.Source: §I · p. 1
- The framework combines distributed ISAC and HDC for risk-aware passive-entity sensing, with ray-tracing and ROS-based validation on typical robot microcontroller hardware.Source: §I · p. 2
II. SYSTEM MODEL
The system uses synchronized, beamformed channel measurements from multiple receivers to classify passive targets as safe or dangerous. The primer describes HDC representations and the operations used to process them.
II. SYSTEM MODEL
- A synchronized single-antenna transmitter and multiple multi-antenna receivers collect channel impulse responses across beam directions for passive targets on the factory floor.Source: §II · p. 2
- The proof-of-concept classifies targets as safe or dangerous and supplies danger-zone information for ROS path adaptation, while remaining an auxiliary rather than standalone safety-critical system.Localization, trajectory estimation, and multi-class scene understanding are left for future work.Source: §II · p. 2
III. A PRIMER ON HYPERDIMENSIONAL COMPUTING
- HDC represents data as high-dimensional, noise-tolerant hypervectors whose distributed information and same-dimensional operations support scalable similarity-based processing.Source: §III · p. 2
- Bundling aggregates hypervectors into a composite similar to its inputs, whereas binding forms associations and produces a result approximately orthogonal to both operands.Source: §III · p. 2
- Permutation circularly shifts a hypervector, helping distinguish identical elements that occur in different orders.Source: §III · p. 3
IV. HYPERDIMENSIONAL COMPUTING FRAMEWORK
The framework encodes channel measurements into hypervectors for classification. It distinguishes continuous and binary MAP representations and constructs dictionaries for beam angles, delays, powers, and receiver identities.
IV. HYPERDIMENSIONAL COMPUTING FRAMEWORK
- Each sample is a channel impulse response tied to a receiver, beam direction, and time slot; MAP maps a real-valued input into a high-dimensional vector for classification.Source: §IV · p. 3
- Continuous MAP retains real-valued accumulated hypervectors and uses cosine similarity, while binary MAP applies a sign-based binarization for lower-complexity storage and processing.Source: §IV · p. 3
- The framework builds separate hypervector dictionaries for azimuth and elevation beam angles, delays, powers, and receiver IDs.Source: §IV · p. 3
A. Group-level encoding
The encoding groups CIR measurements by target position, represents each row’s multipath taps together with its receiver and beam configuration, then combines rows into a channel signature for that position.
A. Group-level encoding
- Each group contains all receiver measurements for one target position, with each row recording a CIR at a fixed receiver, beam configuration, and time slot and labelled safe or danger.Source: A. Group-level encoding · p. 3
- Each row’s CIR contains T delay–power taps representing multipath components, including LoS and nLoS paths.Source: A. Group-level encoding · p. 3
- The row encoding binds each tap into a hypervector, permutes it by tap number, and binds the result with the beam’s azimuth and elevation and the receiver ID.Source: A. Group-level encoding · p. 3
- Rows in a target-position group span the full 3D beamsweep, and their combined signature represents the channel at that target location.Source: A. Group-level encoding · p. 3
B. Training
Training builds one safe and one danger class hypervector by accumulating group representations, with class-count weighting to account for unequal class sizes.
B. Training
- Group hypervectors are weighted by inverse class count because safe and danger samples may differ substantially in number.Source: B. Training · p. 3
- A class hypervector is formed by bundling group-level hypervectors with the representative hypervector for their known safe or danger label.Source: B. Training · p. 3
- The warehouse layout illustration marks receiver beams as blue dots and danger and safe zones in red and green.Source: B. Training · p. 4, Figure 5 · p. 4
C. Iterative Retraining
Iterative retraining seeks better class separation by updating prototypes on misclassified samples, then normalizing them and stopping at convergence or the epoch limit.
C. Iterative Retraining
- Training shuffles samples and compares each sample’s cosine similarity with current class prototypes to estimate its label.Source: C. Iterative Retraining · p. 4
- When a sample is misclassified, the class hypervector prototypes are updated using its true and estimated labels.Source: C. Iterative Retraining · p. 4
- Prototypes are renormalized each epoch; training tracks the minimum training error and stops early at zero errors or after the maximum epochs, supporting rapid adaptation to environmental changes.Source: C. Iterative Retraining · p. 4
D. Inference
Inference fuses CIR measurements from all receivers during concurrent 3D beam sweeps into one hypervector, then assigns the unknown target location the class with the highest similarity.
D. Inference
- The fusion module group-encodes concurrent 3D beam-sweep CIRs from all receivers, then labels the unknown location by the most similar safe or danger class hypervector.Source: D. Inference · p. 4
V. NUMERICAL RESULTS AND DISCUSSION
Joint processing of distributed receiver measurements improves danger-zone classification over either receiver alone, while similarity maps show cleaner spatial separation. The evaluation uses ray-traced warehouse measurements across safe and danger regions.
V. NUMERICAL RESULTS AND DISCUSSION
- The warehouse ray-tracing setup uses a moving forklift, one transmitter, two receivers, and 30 GHz operation.The floor is divided into danger and safe zones, with CIRs collected from receive-beam sweeps over a grid of locations.Source: §V · p. 4
- CIR power uses 109 quantile bins, while multimodal tap delays use 267 uniform bins to preserve a constant nanoseconds-per-bin ratio.The delay binning is intended to reflect physical proximity and temporal scale in hypervector similarity.Source: §V · p. 4
- 88.98% mean accuracy and 84.94% danger F1-score for Joint RXs, versus 84.06% and 80.51% for RX 1 and 82.74% danger F1 for RX 2.Joint processing also reduces accuracy standard deviation to 0.0319, compared with 0.0518 for RX 1.Source: §V · p. 4, Figure 6 · p. 4
- Joint RXs show a sharper danger-region boundary and uniform safe-region bias, mitigating the localized red false-alarm patches seen with individual receivers.The maps plot median Danger-minus-Safe similarity across the spatial grid; red and blue indicate danger and safe class dominance.Source: §V · p. 5, Figure 7 · p. 5
A. Benchmarking
The section compares HDC encodings with a resource-heavy DNN baseline using raw CIR samples. Continuous MAP achieves the highest reported accuracy, while binary encoding retains an accuracy advantage and is optimized for faster inference.
A. Benchmarking
- 88.98%(±3.19%) mean accuracy for continuous MAP, versus 77.39% for the DNN and 82.21% for binary encoding.The comparison is in the joint scenario; the DNN uses eight layers and approximately 465,000 trainable parameters.Source: A. Benchmarking · p. 5, Figure 8 · p. 5
- Binary encoding is optimized for ultra-fast inference and achieves 82.21% accuracy, above the DNN baseline’s 77.39%.The paper describes this as accepting a marginal accuracy trade-off for reduced latency.Source: A. Benchmarking · p. 5
- Danger F1-score is 0.8494 for MAP and 0.6642 for the DNN, with the DNN struggling to generalize the Danger zone to test locations.The authors attribute the DNN behavior to overfitting to quantization noise at training coordinates.Source: A. Benchmarking · p. 5
B. MCU Performance
Binary HDC operations show substantial execution-time advantages on an embedded MCU, and ROS2 integration uses inferred occluded-object regions to adjust navigation costs. The simulation illustrates how global costmap changes can influence planning before the robot reaches a blind spot.
B. MCU Performance
- At D = 10,000, binary bundling takes 248 µs versus 4260 µs for MAP on the STM32 MCU.Binding and similarity operations show speedups exceeding 15× and 17×, respectively; Table I summarizes execution times for three HV dimensions.Source: B. MCU Performance · p. 5, Table I · p. 6
C. ROS Simulation
- The ROS2 node represents inferred occluded objects as elevated-cost cells in the global costmap, preserving compatibility with standard global planners.The object region is converted to a 2D point cloud and projected into the map as an occupancy grid.Source: C. ROS Simulation · p. 6
- Global cost adaptation makes the occluded region non-traversable in the illustrated scenario, producing a path with safe clearance before the robot enters the blind spot.The authors contrast this long-horizon planning with local-costmap-only updates, which may prompt avoidance only as the robot approaches the region.Source: C. ROS Simulation · p. 6
VI. CONCLUSIONS
The framework combines distributed devices and HDC to extend perception beyond occlusions, using infrastructure-based channel measurements for robust, scalable inference without explicit channel reconstruction. Ray-tracing simulations report up to 89% detection accuracy in complex indoor environments while maintaining computational requirements on typical robot hardware.
VI. CONCLUSIONS
- The framework combines spatially distributed devices with HDC, encoding multipath-induced channel variations into high-dimensional representations for inference without explicit channel reconstruction.The approach uses infrastructure-based channel measurements and cooperative transmissions to extend perception beyond occlusions.Source: §VI · p. 6
- Fig. 9 shows the resulting costmap for a given navigation goal.Source: Figure 9 · p. 6
- Up to 89% detection accuracy is reported in complex indoor environments, with computational requirements maintained on typical robot hardware.The result comes from ray-tracing simulations.Source: §VI · p. 6