IntelX

Abstract

Autonomous exploration requires a robot to decide where to move next based on what it has observed so far. Classical frontier methods only use the geometry of the occupancy grid and commit to targets that yield little new coverage at the cost of longer traversal. Recent learned methods instead complete the unseen map with a generative model, which synthesizes geometry the planner never consumes and risks hallucinating structure. In this paper, we show that exploration needs only an estimate of how much new area a candidate goal would reveal. We propose IntelX, a framework that learns this estimate with a single lightweight U-Net, regressing coverage gain densely over a robot-frame occupancy image under supervision from a simulation oracle that measures the coverage actually obtained from every reachable nearby goal. The network returns the field over a 20 m window in 8.1 ms, and because its input is an occupancy image it can potentially adapt to other sensor configurations. We show that the same field also exposes topological structure, with junctions appearing as multi-peaked angular mass. The IntelX planner commits to their branches weakest-first, clearing terminating branches before corridors that stay reachable later. Our experiments on three held-out floor-plans and three out-of-distribution corpora show that IntelX reaches the 99% coverage target on large maps over paths more than 20% shorter than any of the eight baselines that also reach it, achieving high area covered per meter and per second efficiencies.

Live Demo

Run IntelX on a floor plan of your own, entirely in your browser. Pick an example, draw walls and rooms, or upload a map (a ROS map, or any image with dark walls on a light floor), then press Start. The page downloads the trained U-Net (26 MB) and runs it on your graphics card through WebGPU, or on the processor where WebGPU is unavailable. The simulator and both planners are a direct port of the released code, with the benchmark's sensor and settings, and nothing you draw or upload leaves your computer.

Note: the demo runs a smaller, compressed copy of the IntelX network, its weights stored at 8 bits instead of 32 to keep the download small. It may be less accurate than the full model evaluated in the paper, and by default it also predicts less often than the paper's runs (every 30 cm, or 60 cm on a processor, instead of 5 cm), so results here can differ from the paper's.

Example map

Map

0.0%Coverage

0 mPath

Predict Every

Predicted Field Planner

Overview

To decide where to go next, a robot does not need to imagine the rooms it has not seen. It needs one number for every goal around it: how much new area it would reveal by standing there. IntelX learns that number with a single lightweight U-Net, and reads the structure of the building straight off the same field.

Map Completion

“What does the unseen space look like?”

Synthesizes geometry with a generative model, paying in parameters, data and latency for detail the planner never consumes.

MapEx · UPEN · IG-Hector

End-to-End Policies

“Where should I go?”

Folds perception and strategy into one network that needs interaction data and gives no account of its decisions.

Reinforcement and imitation learning

IntelX

“How much new space would I gain by standing there?”

A supervised regression target with a simulation oracle. No generative model, and directly actionable.

One U-Net · One field · Two planners

37%Shorter path to 99% coverage than UPEN on Large Office 7
21%Shorter than MapEx on Medium Office 6
34%Shorter than Frontier (AutoX) on Small Office 3
8.1 msPer 20 m coverage-gain field, about 120 FPS

Path to 99% (P99) among the methods that complete each held-out map, 20 trials per map (Table 3).

Video

Method

Local decisions are learned, since the value of a nearby goal turns on geometry too subtle to hand-design. Global decisions are not, since indoor structure is discrete and legible: junctions are read off the same field as multi-peaked angular structure and managed with explicit memory rather than a second network.

IntelX overview. A robot in an unseen 2D environment observes with a 220 degree laser. Its local window is turned into a robot-frame occupancy image, a U-Net predicts a coverage heatmap, the field splits into ranked components and junction branches, and the local and global planners use them, falling back to frontier centroids when stuck. — **Overview.** The robot observes with a 220° laser, focuses on a 20 m window in its own frame, asks the predictor how much more it can see, reads candidate directions and junctions off the field, and plans locally and globally, falling back to frontier centroids when stuck.

Architecture

The IntelX Dataset

A custom 2D simulator ray-casts a 220° laser of 662 beams out to 10 m over 18 floor plans, standardized to 0.05 m per pixel and sealed watertight. A show walks outward from a seeded spawn through 10 seasons. At each season, 157 goals on a 2.5 m ring are replayed as episodes from the frozen state, and the coverage gain Δρ that each one actually yields is recorded along its path. Every season becomes one training pair of a robot-frame occupancy image I_t and its field H_t: 10⁶ pairs from 100,000 shows, collected by up to 720 concurrent simulators.

Colour scale for coverage gain Δρ, from 0 to about 0.05

Shows per training map (Table 1), allocated in proportion to a complexity score κ that balances room count against explorable area, so the large offices that carry junction structure are not starved. Hover a bar for its rooms, area and κ.

Coverage Gain Predictor

Given the robot's local scan image I_t, the network regresses the expected coverage gain Δρ at every pixel of the 20 m window. It is a U-Net with a ResNet-34 encoder at depth 5, adapted for single-channel input and trained from scratch. Four decoder stages of width 256, 128, 64 and 32 with bilinear upsampling and skip connections end in a sigmoid over the 400 × 400 window, 24.4 M parameters in all.

Training uses a weighted MSE (λ = 10⁸), which keeps the gradients of gains as small as 10⁻⁴ clear of the FP16 denormal range. Because the input is an occupancy image, the same predictor can be retrained for other ranges, fields of view and beam counts, confining the sim-to-real gap to how occupancy is rendered.

Predicted fields against ground truth in four situations: at the initial spawn the prediction is bidirectional and correct, partly explored it predicts gain behind the robot, inside an explored room it predicts low coverage, and at a junction the ranking of two directions is reversed. — Predicted Ĥ_t against ground truth H, each on its own scale. (a)–(c) are correct; at the junction in (d) the ranking of two directions is reversed.

Local Planner

The field is smoothed and its mass gathered into a 72-bin angular histogram around the robot. Its peaks split the window into angular components, and when a single component survives, its intensity-weighted centroid becomes the goal. The global maximum sits deep inside a room, and one centroid over the whole field lands almost on the robot, since opposing lobes cancel. The component centroid is the only goal far enough to be worth travelling to and near enough to explore the room completely.

The four candidate goals of Fig. 4(d), at one junction.
Goal	Distance [m]	Δρ
(a) Global Maximum	2.5	0.0168
(b) Weighted Centroid	0.3	0.0085
(c) Component Centroid (IntelX)	1.5	0.0168
(d) Frontier Centroid	6.6	0.0168

Local planner ablation. Four candidate goals at one junction: the global maxima of the field, the weighted centroid of the whole field, the component centroid used by IntelX, and the frontier centroid. Right: the local planner alone stalls at 89.9 percent coverage while the predicted field collapses. — (a)–(d) Four goals at one junction; the component centroid μ_{C_g} matches the best gain at the shortest distance. (e)–(h) Alone, the local planner stalls at 89.9%: the field collapses from max Δρ = 2.1 × 10⁻¹ to 3.8 × 10⁻⁴, and recovery needs a target outside the window, which only the global planner can supply.

Global Planner

The global planner turns multi-peaked structure in the field into persistent, rankable branches. A junction registers only when at least 5 of 11 poses, the robot and 10 virtual poses on a 0.5 m circle, see more than one component. A branch's strength predicts what lies beyond it: a weak branch ends in a room or a dead end, while a strong one opens onto a corridor. Clearing the weak branch first costs a short detour now instead of a long return later.

IntelX in Action

The predicted field in the world frame, its ranked components, the junctions and their branches, and the frontier fallback, as the robot sees them.

Results

Each trial draws one spawn by seeded rejection sampling and every method runs from that pose, so a difference between methods is attributable to the policy rather than the start. Below, the trials of the paper's Fig. 6 and of the supplementary video are replayed for all nine methods, synchronized by distance travelled: the three held-out KTH floor plans, seven out-of-distribution maps from HouseExpo, Modified Swiss Dwellings and AWS RoboMaker, and the fourth floor of the ECSS building at UT Dallas, a map built with SLAM from a real robot's laser scans. Pick a map, then play or scrub; hovering over a method highlights its coverage curve.

KTH Test

Out of Distribution

Real SLAM Map

Distance travelled 0 m

Held-Out Floor Plans

20 trials per map; P99 averages only the runs that reach ρ^⋆ = 99%. Among the methods that complete a map, IntelX takes the shortest path on all three.

Out-of-Distribution Corpora

10 trials per map. Primed path lengths are per-map ratios to the best method, so 1.00 means shortest.

MapEx is the closest competitor, and the comparison is not uniform over a run. Below 80% coverage it leads IntelX on roughly half the runs, since predicted completions pay off while large unseen regions remain. Above 90%, its curve flattens as the remaining gain falls below its prediction resolution, and IntelX overtakes. The same effect favors MapEx on smaller maps, where the floor plan is close to its prediction horizon and the late phase is short.

Ablations

Distance travelled 0 m

Weakest First, With the Learned Field. On Medium Office 6, IntelX reaches 99% in 516 m. Committing to the strongest branch first needs 679 m, 32% more, and replacing the learned field with geometric information gain stops at 98.2% after 1,060 m.

Distance travelled 0 m

Local Alone Stalls. On Small Office 1 the local planner alone stalls at 91.1% after 88 m, planning goal after goal without moving. With junction memory and the frontier fallback, IntelX reaches 99% in 124 m.

BibTeX

Coming soon. The BibTeX for the paper will be posted here once it is on arXiv.