Preference Adaptation in Collaborative and Assistive Robotics

44 concepts40 papers1 proposal

When autonomous robots transition from controlled factory floors to human-centric spaces like homes and care facilities, they encounter a fundamental challenge: human heterogeneity. Traditional robotics excels at optimizing objective, well-defined metrics—such as minimizing path length, maximizing mechanical efficiency, or ensuring static stability. However, human environments are governed by highly subjective, context-dependent preferences. What constitutes an "organized room," a "natural walking gait," or a "polite interaction" varies dramatically from person to person.

Because there is no universal ground truth for these tasks, a one-size-fits-all control policy inevitably fails. To achieve seamless human-robot interaction (HRI), modern systems must treat human preferences not as fixed constraints, but as latent variables that must be inferred, adapted to, and tracked over time.

Inferring Preferences from Environmental Context

The most direct way to discover how a user wants a task performed is to ask them. However, constantly interrogating a user creates a high cognitive burden and a frustrating user experience. To bypass this, recent frameworks leverage the idea that the environment itself acts as a static footprint of a user's dynamic preferences.

For instance, in domestic tidying tasks, a robot must balance universal common sense (e.g., plates belong in the kitchen) with highly idiosyncratic habits (e.g., a specific book belongs on the nightstand). To capture this nuance without manual configuration, Wenhao Li et al. developed a framework that utilizes Large Multimodal Models (LMMs) to transform standard scene graphs into Affordance-Enhanced Graphs (AEGs) in LLM-enhanced Scene Graph Learning for Household Rearrangement. By analyzing the existing, undisturbed state of a home, the system mines localized user preferences implicitly.

Similarly, the authors of PARSEC: Preference Adaptation for Robotic Object Rearrangement from Scene Context introduced ContextSortLM, which represents spatial preferences using structured JSON outputs generated directly from scene contexts. This allows the robot to infer personalized placement logic that generalizes to entirely new objects and environments.

[Raw Scene Graph] ──> [LMM / ContextSortLM] ──> [Affordance-Enhanced Graph] ──> [Personalized Trajectory]
  (Current State)                                 (Latent Preferences Mapped)        (Tailored Rearrangement)

This contextual inference extends to understanding underspecified language. When a user asks a robot to "clear my usual workspace," the command cannot be executed using generic semantic maps. To bridge this gap, Taeyoon Kwon et al. proposed MEMENTO in Embodied Agents Meet Personalization: Investigating Challenges and Solutions Through the Lens of Memory Utilization. MEMENTO utilizes a hierarchical knowledge graph memory module that separates raw interaction history from long-term personal facts, allowing the robot to ground ambiguous commands in the user's specific history.

Perceptual alignment is critical because even basic descriptive words are subjective; as Natalie Parde et al. demonstrated in Grounding the Meaning of Words through Vision and Interactive Gameplay, a robot must actively align its internal visual models with the specific perceptual biases of its human partner through interactive dialogue to ensure true shared understanding.

Optimizing High-Dimensional Physical Interactions

When personalization shifts from high-level task planning to low-level physical dynamics—such as tuning a robotic exoskeleton or executing a fluid object handover—the optimization space becomes high-dimensional and continuous. Here, the core bottleneck is human evaluation fatigue: humans are notoriously poor at providing absolute numerical ratings for physical comfort, and they cannot endure hundreds of experimental trials.

To solve this, Maegan Tucker et al. developed the LINECOSPAR framework in Human Preference-Based Learning for High-dimensional Optimization of Exoskeleton Walking Gaits. The intuition behind LINECOSPAR is to break a daunting, high-dimensional gait optimization problem into a sequence of simple, one-dimensional line searches. Instead of asking a user to rate a gait on a scale from 1 to 10, the system presents two sequential gaits and asks a simple binary question: Which one felt better? By leveraging these pairwise comparisons, the system maps the user's underlying utility function—balancing factors like dynamic movement versus perceived safety—and identifies the optimal personalized gait in fewer than 30 trials.

High-Dimensional Gait Space ──(Decomposition)──> Sequential 1D Line Searches ──(Pairwise Choices)──> Rapid Convergence

In close-contact scenarios, physical receptivity is highly person-specific. As Ali Reza Manzoori et al. showed in Evaluation of controllers for augmentative hip exoskeletons and their effects on metabolic cost of walking: explicit versus implicit synchronization, different individuals respond radically differently to explicit synchronization (calculating gait cycles) versus implicit synchronization (reacting to foot pressure). This variance necessitates real-time, adaptive control loops.

For instance, in robot-to-human object handovers, Andras Gabor Kupcsik et al. utilized contextual policy search combined with a Bayesian reward estimation in Learning Dynamic Robot-to-Human Object Handover from Human Feedback to dynamically adjust arm compliance and grip based on immediate human feedback.

When the physical interaction is even more intimate, such as assistive feeding, pre-programmed velocities fail due to user anxiety or physical spasms. To counter this, Rajat Kumar Jenamani et al. developed a multimodal controller in Feel the Bite: Robot-Assisted Inside-Mouth Bite Transfer using Robust Mouth Perception and Physical Interaction-Aware Control that differentiates between intentional bites and accidental collisions, adapting its velocity control to match the real-time physiological comfort levels of the individual.

Dynamic Task Planning and Division of Labor

True collaboration requires more than just executing a trajectory safely; it requires an algorithmic understanding of who should do what, and when. Even when a human-robot team has a clear objective goal, the optimal division of labor remains ambiguous because humans hold unobserved preferences regarding their contribution roles.

As Michelle D Zhao et al. argued in Learning Human Contribution Preferences in Collaborative Human-Robot Tasks, failing to align with these latent role preferences leads to sub-optimal team efficiency and user dissatisfaction. To dynamically adapt task sequences, Vidhi Jain et al. introduced the Transformer Task Planner (TTP) in Transformers Are Adaptable Task Planners. TTP treats physical objects in a scene as tokens within a sequence, allowing the model to ingest a single demonstration from a user and immediately replicate their specific stylistic constraints and sequencing preferences in entirely novel environments.

[Object Tokens in Scene] ──> [Transformer Task Planner] ──> [Sequence Aligned to User Style]

Furthermore, communication styles must be personalized. In real-time collaborative scenarios, some users prefer explicit verbal updates, while others find frequent robot-initiated communication intrusive. To manage this variance, Shipeng Liu et al. designed MonTA (Monitor-then-Adapt) in Benchmark Real-time Adaptation and Communication Capabilities of Embodied Agent in Collaborative Scenarios. MonTA splits the robot's cognitive architecture into a fast, embedding-based monitor (System 1) and a deliberate LLM planner (System 2), allowing the robot to tailor its communication frequency and task delegation style to the individual user's expectations.

This balancing act is particularly visible during human-induced conflicts, such as when a human unintentionally blocks a robot's path. Dongping Li et al. introduced the COMMET framework in COMMET: A System for Human-Induced Conflicts in Mobile Manipulation of Everyday Tasks, which uses fine-tuned Vision-Language Models to reason over situational urgency and user "vibe," deciding whether the robot should politely wait, reroute, or explicitly ask for assistance.

This capacity to shift interaction modalities—from robot-led to human-led or fully adaptive—underpins the flexible supervision architecture proposed by Michelangelo Fiore et al. in On Planning and Task Achievement Modalities for Human-Robot Collaboration.

Sample Efficiency and Generalization

A personalized robotic system is impractical if it requires a massive, time-consuming retraining session for every new user or task. The frontier of preference adaptation research focuses on sample efficiency: transferring learned preferences across distinct domains and robustifying policies before they ever interact with a real human.

One powerful approach is to map user behaviors to task-agnostic features. Humans often follow internal heuristics, such as frontloading difficult tasks or minimizing cognitive load. Runyu Guan et al. demonstrated in Towards Transferring Human Preferences from Canonical to Actual Assembly Tasks that by modeling user preferences through these abstract features of physical and mental effort, a robot can observe a user's "work style" during a simple 6-step canonical task and successfully transfer that preference model to a highly complex, 17-step actual assembly task.

[6-Step Canonical Task] ──> Extract Task-Agnostic Features ──> [17-Step Assembly Task]
                              (Physical & Mental Effort)

Alternatively, robots can be pre-trained to handle human variance before deployment. In Robustifying a Policy in Multi-Agent RL with Diverse Cooperative Behaviors and Adversarial Style Sampling for Assistive Tasks, Takayuki Osa et al. addressed the risk of overfitting to a single user profile by using mutual information to autonomously generate a diverse array of simulated human interaction styles. By employing adversarial sampling during multi-agent reinforcement learning, the robot is forced to find a robust policy that can rapidly adapt to unseen, real-world partners.

Complementing this, Hongtao Li et al. developed ICLHF (In-Context Learning from Human Feedback) in In-situ Value-aligned Human-Robot Interactions with Physical Constraints, which pairs the rapid in-context reasoning of LLMs with a physics-aware pose optimizer. This ensures that as the robot adapts to subjective user preferences on-the-fly, its trajectories remain strictly bounded by physical safety laws.

Finally, when the robot's role shifts from an assistant to an instructor, frameworks like COACH, developed by Cunjun Yu et al. in COACH: Cooperative Robot Teaching, use Item Response Theory within a POMDP framework to track student progress. This allows the robot to abandon rigid, objective teaching templates and instead tailor its instruction to the unique physiological constraints and learning pace of the individual.

Ultimately, by moving away from monolithic optimization targets and embracing frameworks that infer, decompose, and generalize human utility functions, modern robotics is closing the gap between rigid automation and truly aligned, personalized human-robot partnerships.

Key Foundational Work

On Planning and Task Achievement Modalities for Human-Robot Collaboration (2014) — contributed 1 concept elements to this direction.
How may I serve you?: a robot companion approaching a seated person in a helping context (2006) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_personalization_in_human_robot_interacti_0356664e,
  title = {personalization in human-robot interaction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/personalization-in-human-robot-interaction/rdr_0356664eaddf4251b0af3d33483fdfcf},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_personalization_in_human_robot_interacti_0356664e_v1,
  title = {personalization in human-robot interaction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/personalization-in-human-robot-interaction/rdr_0356664eaddf4251b0af3d33483fdfcf/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake