TAS: A Transit-Aware Strategy for Embodied Navigation with Non-Stationary Targets

TL;DR

The paper tackles embodied navigation in non-stationary environments by introducing Dynamic Object Maps (DOMs) to model time-varying object locations and Transit-Aware Strategy (TAS) to align agent routes with moving targets. TAS-RL and TAS-LLM augment traditional RL and LLM-based policies with temporal transit information and a sparse interception reward to improve convergence under object mobility. Across a multi-object finding task on DOMs, TAS-enhanced agents achieve substantial gains in success rate and generalization to dynamic settings, demonstrating improved adaptability over static-target baselines. The work provides a new benchmark for non-stationary embodied navigation and paves the way for future sim2real transfer and mixed-transit scenarios, with code and data released for public use.

Abstract

Embodied navigation methods commonly operate in static environments with stationary targets. In this work, we present a new algorithm for navigation in dynamic scenarios with non-stationary targets. Our novel Transit-Aware Strategy (TAS) enriches embodied navigation policies with object path information. TAS improves performance in non-stationary environments by rewarding agents for synchronizing their routes with target routes. To evaluate TAS, we further introduce Dynamic Object Maps (DOMs), a dynamic variant of node-attributed topological graphs with structured object transitions. DOMs are inspired by human habits to simulate realistic object routes on a graph. Our experiments show that on average, TAS improves agent Success Rate (SR) by 21.1 in non-stationary environments, while also generalizing better from static environments by 44.5% when measured by Relative Change in Success (RCS). We qualitatively investigate TAS-agent performance on DOMs and draw various inferences to help better model generalist navigation policies. To the best of our knowledge, ours is the first work that quantifies the adaptability of embodied navigation methods in non-stationary environments. Code and data for our benchmark will be made publicly available.

Figures (4)

• Figure 1: Transit-Aware Strategy (TAS) We introduce TAS to tackle embodied navigation in non-stationary environments. In this figure, an embodied agent is tasked with finding a wallet that changes positions in the environment between 9:00 and 9:20. $P_{\text{wallet}}$ represents the transit route of the wallet. TAS makes agents "transit-aware" by modeling object transit information to augment the agent's navigation policy for temporally-informed decision making.
• Figure 2: Dynamic Object Maps (DOMs): We introduce dynamism to topological graphs via portable targets following human-habit inspired routes. In this representative figure, the watch (in red) moves from node N1 at T=10:00 AM to node N5 at T=7:45 PM, while the wallet (in orange) moves from node N5 at T=10:15 AM to node N1 at T=7:45 PM. We benchmark navigation agents under different object transition scenarios and present a Transit-Aware Strategy (TAS) to improve their performance.
• Figure 3: Object Transition: Portable objects move around the scene at various timesteps in accordance to their natural rooms (Table I in Appendix) and transit scenarios (Table \ref{['tab:placement_cases']}). Here, a mug is placed on a kitchen node $N_{3}$ at timestep $T=53$, but moves to a bedroom node $N_{5}$ at timestep $T=55$ via node $N_{4}$. If an agent reaches the kitchen after $T=53$ (or the bedroom before $T=55$), it would fail to see the mug. Multiple objects could also be at the same node (hat and mug in $N4$).
• Figure 4: Navigation Generalizability on DOMs (RCS): Relative Change in Success (RCS) clipnav measures an agent's adaptability, comparing dynamic to static TG performance (optimal RCS is $0\%$). Positive values indicate better dynamic performance (chance encounters), while negative values reflect poor adaptability. Both our TAS-enhanced agents approach optimal RCS, demonstrating better generalization to non-stationary environments.