UcON: User-Centric Object Navigation (ICRA 2026)

Official repository for the ICRA 2026 paper:
User-Centric Object Navigation: A Benchmark with Integrated User Habits for Personalized Embodied Object Search.

This repository is being prepared for open-sourcing. Currently, it contains the README and basic project information. Code, task resources, and evaluation scripts will be released after cleanup and verification.

• README
• Environment setup
• Task resources
• Evaluation scripts

UcON evaluates an agent’s ability to leverage a User Habit Knowledge Base (UHKB) to locate a target object category in a scene.

• 489 target object categories; ~22,600 natural-language habits.
• Built on a customized OmniGibson-based simulator with 22 initial scenes.
• Each episode provides a large UHKB; only a small fraction is relevant to the current target.

In each task instance:

• The agent is initialized in a habit-shaped scene.
• The goal is to locate a target object category given a UHKB.

• Action space: MoveAhead, RotateLeft, RotateRight, LookUp, LookDown, Open, Done
  • MoveAhead: 0.25m
  • RotateLeft/RotateRight: 90°
  • LookUp/LookDown: 30°
• Open: opens eligible containers within FoV and distance threshold d_open = 1m
• Success: call Done when the target is visible and within d_succ = 1m
• Max episode length: 300 steps

• Success Rate (SR)
• SPL (Success weighted by Path Length)

We recommend using the official OmniGibson Docker image as the base runtime, then applying UcON’s small code patches inside the container.

A machine with an NVIDIA GPU + working driver Docker installed, plus NVIDIA Container Toolkit enabled (so --gpus all works) This repo cloned on the host (example path below)

docker pull stanfordvl/omnigibson:1.0.0

On the host, set two paths:

• UCON_DIR: where you cloned this repo
• OG_DATA_DIR: a persistent directory to store OmniGibson assets/datasets (recommended)

# Example: edit these two lines to your own paths
export UCON_DIR=/path/to/UcON
export OG_DATA_DIR=/path/to/og_data

cd "$UCON_DIR"

docker run --gpus all -it \
  -e OMNIGIBSON_HEADLESS=1 \
  -v "$OG_DATA_DIR":/data/og_data \
  -v "$UCON_DIR":/ucon \
  --name ucon-dev \
  stanfordvl/omnigibson:1.0.0 \
  /bin/bash

• /data/og_data is the in-container path; we mount OG_DATA_DIR from the host so datasets persist across container recreation.
• UCON_DIR is mounted to /ucon for easy patching and running scripts.

In the container:

cd /ucon
bash sync_ucon.sh

This syncs the modified OmniGibson files shipped with UcON into the installed OmniGibson source tree used by the container.

In the container:

cd /omnigibson-src
python scripts/download_datasets.py

Datasets will be downloaded under the mounted path /data/og_data (so they persist even if you recreate the container).

In the container:

cd /ucon
bash scripts/smoke_test.sh

TODO: Release the task resources and generation scripts and document the format.

UcON releases base task resources and generates concrete tasks on-the-fly in a customized simulator environment.

TODO: Provide runnable commands for all baselines and reproduce the main results.

We will provide scripts to evaluate various baselines on UcON.

Citation information (BibTeX / public link) will be provided after the camera-ready / public release.

If you have any suggestions or questions, please feel free to contact us:

• Hongcheng Wang: whc.1999@pku.edu.cn
• Jinyu Zhu: zhujinyu@stu.pku.edu.cn
• Hao Dong: hao.dong@pku.edu.cn