TCNav-OVON-AR

TCNav-OVON-AR supports navigation and approach-pose selection for object interaction. Tasks evaluate whether a robot can reach a pose satisfying target-region visibility, approach-direction and kinematic reachability requirements. Built on HM3D-OVON, version 0.3 uses a public Stretch robot model and contains 3,091 simulation tasks across 36 scenes and 38 object categories.

Each task specifies a scene, start pose, target category, interaction-region requirements and action budget. The evaluation protocol checks whether the reached pose satisfies visibility, distance, heading, approach-side, kinematic reachability and static collision conditions for the same target region and anchor.

中文说明 · Statistics · Selection · Source records · License

Dataset contents

Configuration	Development	Comparison	Contents
full_navigation	400	900	Navigation tasks from full-episode starts
local_navigation	400	300	Local starts with preceding observation history
side_pairs	172	600	Left/right condition rows sharing a parent task
regions	1683	5217	Evaluator-only mesh regions and anchors
verified_pool	682	2409	All verified base tasks
observation_history	1200	900	Past RGB-D reconstruction references and camera poses
profiles	8	28	Scene-associated robot, sensor and protocol configurations

The comparison partition contains 900 full tasks, 300 local tasks and 300 left/right pairs. Each pair has two condition rows. Configurations share parent tasks and must not be summed as independent samples.

The frozen scene partition contains eight development scenes and 28 comparison scenes, with no scene overlap. Comparison scenes come from public validation data. The pool includes 788 original OVON starts and 2,303 newly generated starts; source identifiers and hashes preserve their origins.

Selection uses deterministic hashing and scene round-robin sampling. Local tasks are drawn from selected full tasks. Pair parents are retained in verified_pool, including parents outside the bounded full-task selection. Seven side-coverage extension tasks are recorded separately and do not alter the primary full-task selection. See the selection manifest for the exact rules.

The pool includes 1,321 spatially distinct start-extension tasks (start_diversity_extension), each with fresh action execution and replay evidence. New starts are at least 0.5 m from existing and earlier sampled start positions within the same scene. These tasks reuse the regional annotations and preserve the original full/local/side comparison subsets. verified_pool is the default configuration for browsing all base tasks.

Loading

from datasets import load_dataset

repo = "wenyilllpps/TCNav-OVON-AR"
pool = load_dataset(repo, name="verified_pool")
full = load_dataset(repo, name="full_navigation")
local = load_dataset(repo, name="local_navigation")
pairs = load_dataset(repo, name="side_pairs")

All seven configurations provide development and comparison splits in Parquet format. The repository is public; loading requires no authentication or custom remote-code loader. For a local copy, replace the repository ID with its directory. Check downloaded file integrity with sha256sum -c checksums.sha256.

Policy inputs and evaluation records

The task configurations expose environment initialization, goal requirements and public profiles. Local histories contain observations at or before the local start, with the remaining full-task action budget.

regions and evaluator/ are evaluator-only. They contain instance identities, region anchors, reference terminals, inverse-kinematics records, action witnesses and replay evidence. Navigation policies must use the declared public task/history fields and exclude this evaluator information.

profiles/ defines robot geometry, sensors, actions and terminal predicates. The consumer safety profile specifies a head-tilt interval of [-80, 30] degrees for the fixed navigation posture. The archived generation interval is [-80, 40] degrees. A rejected out-of-range look action leaves the state unchanged, reports collided=true and safety_rejected=true, and costs one action; unsafe initial states are rejected. Published witnesses were checked against the consumer interval.

Observations and assets

Observation records include camera intrinsics, optical transforms, trace indices, reconstruction identifiers and RGB-D hashes. Local histories contain past observations; terminal observations are kept with evaluator evidence. Reconstruction identifiers are not image URLs.

This repository distributes task metadata and verification records. HM3D scene assets, raw RGB-D/semantic images, model weights, robot meshes and simulator execution code are obtained separately. Pixel reconstruction requires the recorded assets and matching renderer/runtime; implementation hashes are listed in source records.

Validation scope

Reference tasks passed independent action replay and joint terminal checks, with additional head-motion, ground-contact and robot self-occlusion audits. The captures render the environment; robot occlusion is checked separately using collision geometry in the navigation posture.

This release evaluates simulated navigation and static terminal geometry with the Stretch reference model. Regions are automatically derived; independent human semantic review is incomplete. Left/right conditions use a fixed geometric regional frame. The records do not establish physical manipulation success or reproduce the paper's method-comparison results. Per-record verification and review fields, together with validation reports, specify the evidence associated with each task.

Sources and license

The dataset builds on HM3D-OVON, the OVON episode distribution, HM3D, HM3DSem and the Stretch URDF.

License scope and attribution are defined in NOTICE.md. Source assets and derived annotations remain subject to the applicable upstream terms.