IntentNav: Learning Spatial-Visual Object Navigation from Human Demonstrations

CoRL 2026

Yuxin Cai1,2,4,* Zongtai Li2,* Maonan Wang3 Muyi Bao2 Haokun Zhu2 Ruofei Bai1,4

Ding Zhao2 Zirui Li1 Wenshan Wang2 Wei-Yun Yau4 Ji Zhang2 Chen Lv1,†

1 Nanyang Technological University 2 Carnegie Mellon University

3 The Chinese University of Hong Kong 4 A*STAR Institute for Infocomm Research (I2R)

* Equal contribution. † Corresponding author.

Paper Code Coming Soon Watch Demos
Key Moments - Click image to jump • Click 🔍 to zoom
Geometry-aware cubicle-by-cubicle search
0:08
Geometry-aware search: the agent probes each small region one by one, human-like — glancing into a cubicle and moving on without exhaustively clearing it.
Looking around to avoid redundant revisits
0:16
Active looking-around behavior tracks what has been seen, avoiding redundant revisits to the same regions.
Target found, walking decisively toward it
0:24
The target is found: exploration ends and the agent commits, walking decisively toward the object.
Overview

Learning human-like object search in a unified BEV decision space.

ObjectNav asks a robot to find an unseen target in an unknown home under partial observability. The core difficulty is deciding where to explore next over a long horizon — long before the target is ever observed.

Why

VLM policies lose either space or visual detail

First-person RGB policies keep appearance but discard spatial structure and act reactively over short horizons; text map summaries keep layout but flatten geometry and fine-grained visual cues. Both drift into repetitive search — turning in place, backtracking, and revisiting the same rooms.

How

Recover search intent from human demonstrations

Frontier-based Human-Intent Labeling looks ahead in low-level human trajectories and labels the frontier that best explains the demonstrator's future search direction — turning raw human actions into high-level BEV waypoint supervision, with no oracle shortest paths.

What

BEV-grounded spatial-visual candidate policy

A VLM compares frontier and target candidates in one BEV decision space, each grounded by its egocentric visual memory and agent-centric geometry, then emits a reserved candidate-ID waypoint that transfers across embodiments.

IntentNav architecture: frontier-based human-intent labeling and a BEV-grounded spatial-visual candidate policy
IntentNav architecture. Left: human rollouts update a persistent BEV map, and Frontier-based Human-Intent Labeling recovers the intended frontier as candidate-level supervision by looking ahead at how frontiers evolve in the future. Right: the VLM fuses the BEV decision space, candidate-aligned egocentric visual memories, and agent-centric geometry, then predicts a reserved candidate-ID token as the next waypoint.
Key Contributions

Three ideas that make search human-like and transferable.

1

Frontier-based Human-Intent Labeling

We replay Habitat-Web human ObjectNav demonstrations and convert their low-level actions into candidate-level waypoint labels. When a target is verified we supervise it directly; otherwise we infer intent from how frontiers evolve, labeling the frontier the demonstrator actually committed to.

  • 23,767 human episodes → 2.36M candidate-level samples
  • No oracle shortest-path supervision
2

BEV-Grounded Spatial-Visual Policy

Exploration and target commitment are unified as selection over BEV candidates. Each candidate carries its first-observed egocentric visual memory and an agent-centric geometry feature, and an Intent-Aligned Objective smooths frontier choices by direction while keeping hard commitment to verified targets.

  • 32 reserved candidate-ID tokens — selection, not free-form text
  • Angular soft target for directionally consistent search
3

State-of-the-Art & Zero-Shot Transfer

Trained only on MP3D demonstrations, IntentNav reaches the best SR and SPL among learning-based methods on MP3D, HM3D-v1, and HM3D-v2. The same checkpoint and candidate-level BEV interface transfer with no VLM fine-tuning across three physical robots.

  • 82.2% SR on HM3D-v2 • 70.5% on HM3D-v1
  • Wheeled, Go2 quadruped, G1 humanoid • ~0.8s / decision
Real-World Demos

One waypoint policy, multiple embodiments.

Wheeled Robot

Executes IntentNav waypoint decisions through a mobile base controller.

Unitree Go2 Quadruped

Uses the same high-level waypoint interface on a legged platform.

Unitree G1 Humanoid

Uses the same high-level waypoint interface on a humanoid platform.

Wheeled Robot

Long-Horizon Indoor Search

Within a relatively short path, the robot sweeps an unknown room almost in a full loop, achieving broad spatial coverage of the space before moving on.

Wheeled Robot

Local Scanning

At an informative viewpoint the robot scans in place, rotating to take in its surroundings and gather visual evidence before deciding where to search next.

Wheeled Robot

Probe, Back Out, Commit

A V-shaped peeking motion — walking into a region, finding nothing, and backing out — before semantic evidence of the target appears and the robot commits toward it.

Wheeled Robot

Semantic Common Sense

The open space clearly holds no instance of the target, so the policy skips it and heads straight to another room where the object is semantically more likely to be found.

Wheeled Robot

Unified Exploration & Commitment

Frontier exploration and target commitment are unified in one BEV decision space, so the robot keeps searching and switches to decisive approach the moment the target is verified.

Wheeled Robot

Cross-Room Object Navigation

Carrying its BEV spatial memory across doorways, the robot searches room to room until the target appears.

Unitree Go2

Zero-Shot Quadruped Transfer

The same MP3D-trained checkpoint drives the Go2 with no VLM fine-tuning — only the candidate-level BEV waypoint interface, while platform-specific local control handles legged locomotion.

Unitree Go2

Compact 2B, Onboard

IntentNav is a compact 2B VLM (InternVL3-2B) that runs on a single onboard laptop GPU at roughly 0.8 s per decision — light enough for real-time legged deployment.

Unitree Go2

Very Long-Horizon Exploration

A single extended episode: the quadruped explores a large space over a very long horizon, sustaining coherent directed search without collapsing into repetitive backtracking.

Unitree Go2

Embodiment-Agnostic Interface

Decoupling high-level candidate selection from low-level control lets the identical waypoint policy run on legged hardware exactly as it does on wheels.

Unitree Go2

What Is a Candidate?

Each decision is a choice among frontier candidates on the boundary of explored and unknown space, plus target candidates from verified detections — all compared in one BEV space.

Unitree Go2

Low-Frequency Querying

Instead of repeated VLM calls, the policy predicts a candidate-ID only as the robot nears its waypoint, so 2B inference overlaps with motion and never pauses the robot.

Human / Unitree G1

Zero-Shot Humanoid Transfer

The hardest embodiment gap: the same MP3D-trained checkpoint drives a bipedal humanoid with no VLM fine-tuning, through the same candidate-level BEV waypoint interface.

Human / Unitree G1

Decisive Target Commitment

Once the target is verified, the humanoid stops exploring and walks straight to the object — the high-level policy is unchanged while the body controller executes each waypoint.

Zoomed key moment
Add an image to this key moment to enable zoom preview.
Benchmark Results

Detailed comparison across training regimes and datasets.

Results are reported on MP3D-Val, HM3D (v0.1), and HM3D (v0.2) for fair comparison with prior work.

Method Training Regime MP3D-Val HM3D (v0.1) HM3D (v0.2)
SR (%) ↑ SPL (%) ↑ SR (%) ↑ SPL (%) ↑ SR (%) ↑ SPL (%) ↑
Training-free methods
VLFM Training-free 36.417.5 52.530.4 63.532.5
SG-Nav-GPT Training-free 40.216.0 54.024.9 49.625.5
OpenFMNav Training-free 37.215.7 54.924.4 --
TriHelper Training-free -- 56.525.3 --
BeliefMapNav Training-free 37.317.6 61.430.4 --
ApexNav Training-free 39.217.8 59.633.0 76.238.0
WMNav Training-free 45.417.2 58.131.2 72.233.3
PIGEON Training-free 41.614.4 57.932.3 79.236.8
STRIVE Training-free 52.323.1 62.934.2 79.638.7
SysNav Training-free 50.718.1 63.730.5 80.837.2
Learning-based methods
DD-PPO RL 21.68.7 -- 27.914.2
Habitat-Web Supervised 14.6*4.94* -- --
ZSON Supervised 15.34.8 25.512.6 --
Navid Supervised -- 32.521.6 --
PixNav Supervised -- 37.920.5 --
ESC Supervised 28.714.2 39.222.3 --
PONI Supervised 31.812.1 -- --
SemExp Supervised 36.014.4 -- --
CompassNav Supervised 42.017.5 56.627.6 --
Generic policies
UniGoal Generic 41.016.4 -- 54.525.1
Uni-Navid Generic -- -- 73.737.1
OmniNav Generic -- -- 56.130.0
Ours
IntentNav (Ours) Supervised 53.823.1 70.534.6 82.238.5