Interpreting Context-Aware Human Preferences for Multi-Objective Robot Navigation
Authors: Tharun Sethuraman, Subham Agrawal, Nils Dengler, Jorge de Heuvel, Teena Hassan, Maren Bennewitz
Organizations: Hochschule Bonn-Rhein-Sieg, Germany · University of Bonn, Germany · Lamarr Institute for Machine Learning and Artificial Intelligence, Germany
Robots operating in human-shared environments must not only achieve task-level navigation objectives such as safety and efficiency, but also adapt their behavior to human preferences. However, as human preferences are typically expressed in natural language and depend on environmental context, it is difficult to directly integrate them into low-level robot control policies. In this work, we present a pipeline that enables robots to understand and apply context-dependent navigation preferences by combining foundational models with a Multi-Objective Reinforcement Learning (MORL) navigation policy. Thus, our approach integrates high-level semantic reasoning with low-level motion control. A Vision-Language Model (VLM) extracts structured environmental context from onboard visual observations, while Large Language Models (LLM) convert natural language user feedback into interpretable, context-dependent behavioral rules stored in a persistent but updatable rule memory. A preference translation module then maps contextual information and stored rules into numerical preference vectors that parameterize a pretrained MORL policy for real-time navigation adaptation. We evaluate the proposed framework through quantitative component-level evaluations, a user study, and real-world robot deployments in various indoor environments. Our results demonstrate that the system reliably captures user intent, generates consistent preference vectors, and enables controllable behavior adaptation across diverse contexts. Overall, the proposed pipeline improves the adaptability, transparency, and usability of robots operating in shared human environments, while maintaining safe and responsive real-time control.
Figures & tables
Figure 1: Preference-conditioned navigation behavior in a human-shared supermarket environment. Given a natural language user instruction requesting increased personal clearance, the robot adapts its trajectory to maintain a larger distance from nearby humans while progressing toward the goal location. As can be seen from the baseline shortest-path trajectory (orange) and the preference-adapted trajectory (green), the expressed preference influences the navigation behavior.
Figure 2: Overview of the proposed preference-aware navigation architecture. The reasoning module integrates VLM-based context prediction , LLM-based rule updating , and preference translation to convert visual context and natural language user feedback into structured rules and preference vectors usable by a MORL agent for online low-level navigation control. The preference vectors condition the navigation policy, enabling the robot to adapt its behavior to context-dependent user preferences while providing interpretable explanations of its decisions.
Figure 3: Example output of the context predictor. Given an input RGB scene image and a structured prompt, the VLM-based context predictor extracts semantic and spatial environmental information, including room type, detected objects, approximate object distances, human presence, lighting conditions, and a natural language scene description. This structured context representation is used as grounding input for downstream rule generation and preference translation modules.
Figure 4: Example rule generation for the context-aware preference pipeline. Given an input scene image and a natural language user preference, the VLM-based context predictor extracts structured environmental context, which is combined with user input by the LLM-based rule updater to generate a context-dependent navigation rule and explanation while preserving baseline safety objectives.
Figure 5: Example output of the preference translator. Given the predicted scene context (living room, no humans present) and the stored rule set, the LLM-based preference translator selects relevant context-based rules and generates a four-dimensional MORL preference vector. The resulting vector modulates navigation efficiency, obstacle distance, human distance, and velocity, while also providing the applied rule and a natural language explanation to ensure transparency in preference-conditioned navigation behavior.
Model
Room Type Acc. (%)
Object Acc. (%)
Mean Inf. Time (s)
Gemini-2.0-flash
98.6
53.6
2.2
Gemini-1.5-flash
98.6
48.0
2.9
Pixtral-large-2411
98.6
54.2
9.2
Pixtral-12B
97.1
47.2
4.2
GPT-4o-Mini
98.6
42.2
5.0
GPT-4o
98.6
44.1
7.0
Table 1: Performance comparison of VLMs for context prediction.
Pref. Obj.
Mistral L. 2.1
GPT-4o
GPT-4o-mini
Gemini-2.0-fl
cts
disc
cts
disc
cts
disc
cts
disc
λeffic
0.14
0.27
0.22
0.28
0.35
0.34
0.38
–
λOdist
0.14
0.12
0.12
0.15
0.21
0.15
0.22
–
λDdist
0.04
0.04
0.07
0.04
0.15
0.61
0.81
–
λvelocity
0.12
0.11
0.23
0.22
0.26
0.29
0.20
–
Mean
0.10
0.12
0.14
0.15
0.19
0.29
0.44
–
Table 2: Preference prediction error across evaluated LLMs for both continuous (cts) and discretized (disc) preference vector outputs.
Figure 6: Real-world evaluation environments used for validating the proposed preference-aware navigation pipeline: (a) an office workspace, (b) a home kitchen environment, and (c) a supermarket grocery store. These environments were selected to capture varying spatial constraints, object densities, and human-interaction requirements.
Group
Velocity (m/s)
Human Dist. (m)
Traj. Length (m)
λBase
0.23±0.04
1.01±0.09
2.82±0.03
λHdist
0.17±0.01
1.25±0.03
2.66±0.08
λvelocity
0.09±0.01
1.07±0.07
2.83±0.03
Table 3: Results of the real robot experiments for preference-based navigation averaged over five runs. For each environment, two preference-vector-based navigation behaviors were evaluated. First, the preference to maximize distance from nearby humans ( λHdist ) is activated in all settings. Additionally, environment-specific preferences were applied: reducing velocity when navigating near humans ( λvelocity ) in the office environment, maximizing navigation efficiency or shortest-path behavior ( λeffic ) in the home (kitchen) environment, and maximizing distance from fragile glass bottles ( λOdist ) in the supermarket environment. Note that no human was present in the supermarket during the glass bottle experiments, and vice versa. Therefore, no values are reported for those cases. The results show consistent alignment between the activated preference dimensions and the resulting behavioral adaptations across environments.
Robotic systems that assist humans should be capable of adapting their behaviors to individual user preferences. For instance, users may want a robot arm to adjust the amount of force it applies while folding their laundry or cleaning furniture. Natural language provides an intuitive way for humans to communicate such preferences. Recent progress in language-conditioned robot policies has shown that robots can successfully use language prompts to determine what task to perform. However, extending the same approach to realize how the task should be performed requires detailed labels describing the preferences or styles of trajectories in the task data. Not only is collecting such annotations challenging, but conditioning directly on these labels may also fail to provide fine-grained control over a continuous range of behaviors. For example, it can be difficult to convey the exact force that a robot must apply through abstract instructions like "apply a bit more pressure than before". Therefore, in this work, we propose using language to reason over preferred behaviors instead of directly generating them. We first learn a structured latent representation that organizes user preferences according to differences in the corresponding trajectories. Then, given a preference prompt, we use a foundation model to interpret this latent space and choose a value that produces the desired behavior. Through both simulation and real-world experiments, we show that selecting robot behaviors from an intuitively structured latent space enables more precise adaptation to user preferences while requiring significantly fewer preference labels than language-conditioned policies.
Kevin Robledo, Matías I. Torres Galaz, Kumar Dixhant Rai +3
Department of Computer Science California State University, Northridge · Department of Mechanical Engineering California State University, Northridge
Blind and visually impaired individuals face uncertainty when navigating unfamiliar areas. Whether a cane or a guide dog is used for navigation, this may prove to be challenging. To mitigate this challenge, we present the Context-Aware Navigation Explanation framework that explains guide dog movements for obstacle avoidance and navigation using a large language model. This system integrates perceptual context with planned future trajectories to generate natural-language explanations of robot movement. We deploy the system on a Unitree Go2 platform in indoor navigation scenarios and evaluate how different explanation context configurations affect user trust and explanation satisfaction. Results show a strong positive relationship between explanation satisfaction and trust, and suggest that combining obstacle information with future trajectory leads to the most effective explanations. These findings highlight the potential of context-aware language explanations to improve transparency and trust in assistive robot navigation.
As mobile robots become more integrated into everyday human environments, social robot navigation is becoming essential for ensuring human comfort, safety, and trust. While reinforcement learning (RL) navigation policies provide the fast inference and reactive behavior necessary for real-time deployment, they still lack flexible semantic reasoning capabilities and often fail to generalize to complex social scenarios. Recent approaches have increasingly turned to vision-language models (VLMs) in place of RL policies to improve semantic and social reasoning in robot navigation. Nevertheless, their high computational cost and slow inference remain major barriers to real-time deployment. To overcome these limitations, we introduce HUMA (Hybrid Understanding for Multi-modal social Navigation), a hybrid architecture that dynamically balances the computational efficiency of RL policies with the deep semantic understanding of VLMs. Our approach uses a reactive RL policy to handle low-density, routine navigation tasks, while conditioning it on a post-trained high-level VLM when a human enters sensitive situations, such as the robot's proximity zone. We evaluate HUMA on the Social-MP3D and Social-HM3D benchmarks, where it achieves task success improvements of 20% and 3%, respectively, while significantly reducing personal space violations and human collisions against state-of-the-art baselines. Extensive ablation studies validate each architectural component, and real-world deployment on the Mirokaï mobile robot further demonstrates the practical viability of our approach.
Ali Ahmadi, Hamed Rahimi, Adrien Jacquet Cretides +3
Institut des Syst`emes Intelligents et de Robotique (ISIR) Sorbonne Universit´e France