ATOM: Abstract Transition and Option Modeling - Experience-Informed Hierarchical Navigation

Changxuan Yao; Panagiotis Tsiotras

cs.ROOct 13, 2026

ATOM: Abstract Transition and Option Modeling - Experience-Informed Hierarchical Navigation

Authors: Changxuan Yao, Panagiotis Tsiotras

Organizations: Institute for Robotics and Intelligent Machines Georgia Institute of Technology, Atlanta, GA

Abstract

We propose ATOM (Abstract Transition and Option Modeling), a hierarchical navigation framework that combines higher-level graph planning with reusable lower-level option policies for efficient repeated navigation in large-scale 3D environments represented by hierarchical structures such as 3D scene graphs (3DSGs). ATOM operates on a hierarchical graph representation, implemented using, for instance, the Rooms and Places layers of a 3DSG, where each directed abstract transition is associated with a navigation option. For each transition, ATOM computes and stores a reusable option policy along with its state-dependent cost-to-go. Across repeated navigation tasks, ATOM maintains a smoothed empirical distribution over option initiation states and uses it to progressively refine the corresponding abstract transition cost. Ablation studies show that ATOM reduces cumulative repeated-planning time by 97.5% relative to a monolithic MDP baseline while maintaining trajectories within approximately 7% of a Global Places A* shortest-path reference. Experience-informed refinement further reduces abstract transition-cost estimation error by 10.3%. Physical experiments validate ATOM on an F1Tenth vehicle, successfully completing all five navigation queries with 43% of inter-region option executions reusing previously generated solutions.

Figures & tables

Fig. 1: ATOM framework. Abstract transitions are represented by reusable options, while navigation experience refines transition costs for subsequent high-level planning.

Fig. 2: Hierarchical 3D scene graph of the uHumans2 [ 26 ] office environment generated using Hydra [ 9 ] . The Rooms and Places layers provide the higher- and lower-level graph representations used by ATOM for repeated navigation.

Fig. 3: Representative ATOM query in uHumans2. The Rooms layer shows the abstract route and the Places layer the executed trajectory.

Method	Time (s)	MDP Solves	Solved-State Volume	$\Delta L$ (%)
Global Places A*	$0.144\!\pm\!0.009$	–	–	$0.000\!\pm\!0.000$
Full Graph MDP	$4.955\!\pm\!0.077$	$500.0\!\pm\!0.0$	$246000\!\pm\!0$	$0.000\!\pm\!0.000$
ATOM-NoReuse	$1.945\!\pm\!0.056$	$842.8\!\pm\!18.8$	$77223\!\pm\!1512$	$7.504\!\pm\!0.998$
ATOM-Fixed	$0.102\!\pm\!0.008$	$17.1\!\pm\!0.3$	$1318\!\pm\!9$	$7.504\!\pm\!0.998$
ATOM	$0.122\!\pm\!0.010$	$17.1\!\pm\!0.3$	$1318\!\pm\!9$	$7.303\!\pm\!0.952$

TABLE I: Performance over 10 trials of 500 queries (mean

\pm

sample standard deviation).

Fig. 5: Transition-cost estimation error as option experience accumulates (mean $\pm$ 95% CI across 126 eligible option–trial pairs). ATOM reduces the error from $2.072$ to $1.858$ after 14 executions.

Panel of Figure 7 — Fig. 6: Controlled competing-route experiment. (a) Two alternative routes connect $R_{1}$ and $R_{4}$ . (b) Refinement reverses their estimated cost ordering, causing ATOM to switch routes at query 5; fixed costs retain the original route.

Fig. 7: F1Tenth experimental platform with Intel RealSense D455 and NVIDIA Jetson Orin Nano.

Fig. 8: Representative camera observations collected in the Georgia Tech Van Leer Building. Top: RGB frames from the F1Tenth’s Intel RealSense D455. Bottom: corresponding semantic segmentation overlays used during scene-graph reconstruction.

Fig. 9: Hierarchical 3D scene graph of the second floor of the Georgia Tech Van Leer Building generated using Hydra. The visualization shows the reconstructed environment together with the abstract region nodes and their associated lower-level graph structure used to instantiate ATOM.

Fig. 10: Representative ATOM query on the real-world graph of the second floor of the Georgia Tech Van Leer Building, showing the abstract route (blue) and lower-level trajectory (orange).

Fig. 11: Onboard planning performance on the Van Leer graph over 150 successful start–goal queries. Top left: per-query planning-time distributions for ATOM and Full Graph MDP. Top right: cumulative planning time across successful queries. Bottom left: distribution of the ratio between ATOM trajectory length and the Global Places A* shortest-path reference. Bottom right: cumulative numbers of newly generated and reused inter-region options.

Fig. 12: Physical navigation validation for representative queries. Each panel shows the abstract Room -level route, ATOM-planned Place trajectory, and executed vehicle trajectory. Query 3 includes reuse of a previously generated option.

Robot Navigation

Robot navigation studies how robots perceive, localize, plan, and move safely toward goals in static or changing environments.

See all 554 papers

Visual Navigation Transformer with Pose Attention
Sep 18, 2026Beiming Li, Jaime Romero, Jonathan Diller +2Robot Navigation Attention Mechanisms
GRASP Laboratory, University of Pennsylvania, Philadelphia, PA 19104, USA · Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA 19104, USA
STAG: A Sparse Traversability-Aware Graph Representation from Grid-Based Costmaps for Robotic Navigation
Oct 8, 2026Gabriel Manuel Garcia, Stéphanie Aravecchia, Miguel Angel Olivares-MendezRobot Navigation Graph-Based Planning
University of Luxembourg, Luxembourg · IRL Georgia Tech-CNRS, Metz, France
Learning to Navigate with Minimal Parameters: Decomposing Visual Navigation Through Closed-Form Geometric Interfaces
Jul 13, 2026Edward Beng Wai Tan, Siew-Kei LamVisual Navigation Robot Policy Learning
College of Computing and Data Science, Nanyang Technological University, Singapore.

Sep 18, 2026J/K move · Enter open · S save

Visual Navigation Transformer with Pose Attention

Beiming Li, Jaime Romero, Jonathan Diller, Vijay Kumar+1

GRASP Laboratory, University of Pennsylvania, Philadelphia, PA 19104, USA · Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA 19104, USA

Robot Navigation Attention Mechanisms

Open paper AskPDF arXiv

Learned navigation policies typically consume observations as a temporally ordered history, with positional encodings tying each observation to when it was seen, making it difficult to reuse experience from earlier traversals of an environment. Systems that do reuse such experience usually construct an explicit representation, such as a map or a topological graph, and plan on it. We propose VNT-PA (Visual Navigation Transformer with Pose Attention), a transformer planner whose context is a set of depth keyframes indexed by camera pose. With camera poses as positional encoding, attention depends on the pose differences between keyframes rather than on their temporal order. VNT-PA is trained to imitate a shortest-path planner operating on the ground-truth scene mesh, predicting actions by querying the spatial context with only its current pose and the goal position. On point-goal navigation in HM3D validation scenes, VNT-PA reaches 93.3% success and 90.4% success weighted by path length (SPL), outperforming baselines that encode the same context as a temporal sequence or treat pose as an input feature, in both navigation performance and training efficiency. Because the spatial context is a pose-indexed set, frames from different trajectories can be fused at test time. The planner also degrades more gracefully under localization noise than a conventional baseline which plans on explicit maps. These results show that pose-stamped experience can serve directly as the environment representation for a learned planner, and that making attention depend on pose differences, rather than on temporal order, speeds up training and improves long-horizon navigation.

Abstract

Robot Navigation

Explore similar work

Visual Navigation Transformer with Pose Attention

STAG: A Sparse Traversability-Aware Graph Representation from Grid-Based Costmaps for Robotic Navigation

Learning to Navigate with Minimal Parameters: Decomposing Visual Navigation Through Closed-Form Geometric Interfaces