💻 computer science

Personalized Embodied Navigation for Portable Object Finding

This paper introduces Transit-Aware Planning (TAP) approaches that treat personalized embodied navigation for finding portable objects as a habit learning problem, enabling agents to synchronize with human-moved targets and significantly improving success rates in both simulated and real-world dynamic environments.

Original authors: Vishnu Sashank Dorbala, Bhrij Patel, Amrit Singh Bedi, Dinesh Manocha

Published 2026-04-22

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Vishnu Sashank Dorbala, Bhrij Patel, Amrit Singh Bedi, Dinesh Manocha

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a robot butler named "Robo." In the past, if you asked Robo to find your lost keys, it would assume the keys were sitting exactly where you last left them. It would march straight to the coffee table, look around, and if the keys weren't there, it would get confused and give up.

The Problem: The "Moving Target" Reality
But in real life, things move. You might put your keys on the kitchen counter in the morning, move them to your bag at lunch, and toss them on the nightstand at night. Your phone, wallet, and even your toothbrush follow a daily rhythm.

Current robots are like people with amnesia; they don't remember that objects move. They treat the world like a static museum where nothing ever changes. This paper introduces a new way to teach robots to understand that objects have habits, just like humans do.

The Solution: "Transit-Aware Planning" (TAP)

The authors created a new strategy called Transit-Aware Planning (TAP). Think of TAP as giving the robot a "sixth sense" for movement.

Instead of just asking, "Where is the object right now?" the robot starts asking, "Where has this object been, and where is it likely to go next?"

Here are two creative analogies to explain how this works:

1. The "Bus Stop" vs. The "Chasing Game"

Old Way (The Chasing Game): Imagine you are trying to catch a friend who is running away from you. Every time you run toward where they are, they move to a new spot. You end up running in circles, exhausted, never catching them. This is how old robots work; they chase the object's last known location, but by the time they get there, the object has moved.
New Way (The Bus Stop): With TAP, the robot stops chasing. Instead, it learns the "bus route" of the object. It realizes, "Oh, every day at 9:00 AM, the wallet moves from the bedroom to the kitchen, and at 9:30 AM, it moves to the living room." So, the robot doesn't run after the wallet; it calmly walks to the kitchen and waits there. It intercepts the object's path rather than chasing it.

2. The "Detective with a Habit Book"

Imagine a detective trying to find a missing person.

The Old Detective only looks at the last place the person was seen. If the person isn't there, the detective is stuck.
The TAP Detective has a notebook of habits. They know the person always goes to the gym on Tuesdays and the library on Thursdays. Even if the person isn't at the gym right now, the TAP detective knows to head there because that's where the person usually is at this time of day.

How They Tested It: The "Dynamic Map"

To teach these robots, the researchers couldn't just use normal maps because real houses change. They invented something called Dynamic Object Maps (DOMs).

Think of a standard map as a photograph. It's frozen in time.
A Dynamic Object Map is like a time-lapse video. The rooms (nodes) stay the same, but the objects (the "passengers") hop from room to room on a schedule.

Random Scenario: Objects jump around like popcorn popping (high chaos).
Routine Scenario: Objects move like a train on a track (low chaos, very predictable).

The researchers tested their robots in a virtual world (a video game version of a house) and then in a real lab.

The Results: Why It Matters

The results were impressive:

Better Success: The TAP robots found moving objects 21% more often than the old robots.
Better Adaptability: When the environment changed from "static" (nothing moves) to "dynamic" (things move), the old robots crashed and burned. The TAP robots barely noticed the difference. They generalized 44% better.
The "Toothbrush" Test: In the real-world lab experiment, they hid a toothbrush in a weird place (a workspace, where toothbrushes don't usually belong).
- The Old Robot relied on "common sense" (toothbrushes are in bathrooms) and failed.
- The TAP Robot learned the specific habit of that lab and found the toothbrush in the weird spot because it learned the object's specific route, ignoring the "common sense" rule.

The Big Picture

This paper is a giant leap toward making robots that can actually live with us. Instead of being rigid machines that need everything to stay still, they are becoming observant companions that understand our daily routines.

They aren't just looking for where an object is; they are learning when and how it moves. It's the difference between a robot that is a blind follower and a robot that is a smart, habit-aware assistant.

1. Problem Definition

Current embodied navigation systems (e.g., ObjectNav) predominantly operate under the assumption of static environments where target objects remain stationary throughout an episode. However, real-world environments are dynamic; humans frequently move portable objects (e.g., phones, wallets, toothbrushes) based on habits and routines.

The paper identifies a critical gap: existing Reinforcement Learning (RL) and Large Language Model (LLM) agents fail in these non-stationary scenarios because:

They rely on static graph assumptions.
Standard reward functions (penalizing distance to a target) become noisy and unstable when the target moves, hindering policy convergence.
They lack "temporal awareness" to predict or intercept object transit routes.

The authors formalize this as a Personalized Habit Learning problem, where the agent must learn to synchronize its navigation path with the moving target's path rather than simply chasing a static coordinate.

2. Methodology: Transit-Aware Planning (TAP)

To address these challenges, the authors propose Transit-Aware Planning (TAP), a framework that augments standard navigation policies with temporal knowledge of object movement.

A. Environment Formulation: Dynamic Object Maps (DOMs)

The authors introduce Dynamic Object Maps (DOMs), a dynamic variant of node-attributed topological graphs.

Structure: Nodes represent locations, and edges represent fixed paths. Unlike static graphs, node attributes (objects) evolve over time.
Transit Scenarios ( $\Lambda$ ): Inspired by human habits, three levels of "transit entropy" are defined to govern object movement:
1. Random: Objects move to any room at any time (High Entropy).
2. Semi-Routine: Objects move between fixed rooms but not fixed paths (Medium Entropy).
3. Fully-Routine: Objects move between fixed rooms via fixed paths (Low Entropy).
Goal: The agent must learn the object path sequence $P_o$ to intercept the target.

B. TAP-RL Agent (Reinforcement Learning)

The authors modify standard RL agents (e.g., DD-PPO) to handle non-stationary targets:

Observations: Instead of just local views, the agent receives a snapshot of current object positions $P[t]$ at every timestep.
Reward Function: They replace the standard distance-based reward ( $R_{dist}$ $R_{d i s t}$ ) with a Sparse Interception Reward ( $R_I$ ).
- $R_I$ triggers only when the agent's current node intersects with the target's known path ( $v_c \in P_o$ ).
- This incentivizes the agent to synchronize its route with the target's transit path rather than greedily minimizing distance to a moving target.

C. TAP-LLM Agent (Large Language Models)

For zero-shot LLM-based agents, the authors modify the system prompt to include temporal context:

Context Injection: At each timestep, the agent is fed a history of actions, observed objects, and the current time step.
Mechanism: This allows the LLM to use commonsense reasoning combined with observed transit patterns to predict where an object is likely to be next, rather than relying solely on static scene priors.

3. Key Contributions

Transit-Aware Planning (TAP): A novel strategy to adapt both RL and LLM navigation policies for non-stationary targets by integrating temporal object path information.
Dynamic Object Maps (DOMs): A benchmark framework that transforms static topological graphs (like Matterport3D) into dynamic environments with structured object transitions, simulating human habits.
Generalizability Benchmark: Evaluation of agents across static and dynamic scenarios, introducing metrics to measure adaptability to non-stationary targets.
Real-to-Sim Pipeline: A method to construct DOMs from real-world scene graphs, allowing researchers to simulate and test personalized navigation in physical environments.

4. Experimental Results

The authors evaluated TAP agents on the Matterport3D (MP3D) simulator and in a real-world lab setting.

Simulation Results (MP3D)

Success Rate (SR): TAP agents significantly outperformed vanilla agents.
- TAP-RL: Improved success by 21.1% over vanilla RL in finding non-stationary targets.
- TAP-LLM: Showed consistent improvements across all transit scenarios.
Generalizability (RCS): Measured by Relative Change in Success (comparing static vs. dynamic performance).
- TAP agents achieved a 44.5% better generalization score compared to non-TAP counterparts.
- Vanilla RL agents suffered a massive performance drop (up to -80% RCS) in dynamic environments, whereas TAP agents maintained near-optimal performance (RCS close to 0%).
Ablation: Providing full path information (vs. just current positions) was crucial for RL planning, confirming the need for map-based planning in dynamic settings.

Real-World Results

Setup: A lab environment was mapped to a topological graph with 4 portable targets (mug, laptop, remote, toothbrush).
Performance: TAP-LLM achieved an average success rate of 65-80% in routine scenarios, compared to 37-57% for vanilla LLM.
Non-Commonsense Targets: TAP agents excelled at finding "out-of-place" objects (e.g., a toothbrush in a workspace) by learning specific transit routines, whereas vanilla agents relied on commonsense priors and failed.
Efficiency: TAP agents reduced the "Time-To-Find" (TTF) significantly over time as they learned the object's habit.

5. Significance and Impact

Paradigm Shift: The work challenges the "static graph, stationary target" assumption that has dominated embodied navigation literature for years.
Personalized Assistance: By learning human habits, agents can provide truly personalized assistance, finding items even when they are in unexpected locations, which is critical for real-world deployment in homes and offices.
Robustness: TAP demonstrates that agents can generalize from static training environments to highly dynamic real-world scenarios, a major hurdle in robotics.
Open Science: The authors commit to releasing code, data, and the DOM benchmark to foster further research in dynamic embodied navigation.

In conclusion, this paper establishes that temporal awareness and habit learning are essential for embodied agents to succeed in the real world, where objects are rarely stationary. The proposed TAP framework effectively bridges the gap between static simulation and dynamic reality.