💻 computer science

Mutual Adaptation in Human-Robot Co-Transportation with Human Preference Uncertainty

This paper proposes a unified framework for human-robot co-transportation that addresses human preference uncertainty and adaptation balance by modeling probabilistic human choices, implementing a time-varying stubbornness measure for dynamic leader-follower switching, and utilizing pose optimization to mitigate uncertain behaviors, ultimately enhancing task performance as validated by human studies and simulations.

Original authors: Al Jaber Mahmud, Weizi Li, Xuan Wang

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Al Jaber Mahmud, Weizi Li, Xuan Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ 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 and a robot are carrying a heavy, awkward table through a house full of furniture. You both need to get from the living room to the bedroom without bumping into anything. The problem? You and the robot might not agree on the best route. The robot might think, "That narrow hallway is the fastest way," while you think, "No way, that's too tight and risky; let's take the long way around."

This paper is about teaching the robot how to work with you without you having to speak a single word. It's about mutual adaptation: the robot learning to read your mind, and you and the robot learning to switch roles depending on the situation.

Here is how the researchers solved this, broken down into three simple ideas:

1. The Robot Doesn't Know Your "Brain Settings" (The Uncertainty Problem)

Usually, robots try to guess what you want by picking one specific set of rules (e.g., "This human always hates narrow spaces"). But humans are messy. Sometimes you are brave and take risks; other times you are cautious. Even the same person might act differently on a Tuesday than on a Friday.

The Solution: Instead of guessing one rule for you, the robot creates a "cloud of possibilities."

The Analogy: Imagine the robot doesn't just guess your favorite flavor of ice cream. Instead, it imagines a whole freezer full of flavors you might like, with some flavors being more likely than others.
How it works: The robot calculates the probability of you choosing Path A, Path B, or Path C based on a wide range of possible "risk-tolerance" and "distance-sensitivity" settings. It doesn't assume it knows you perfectly; it assumes it knows a range of who you might be.

2. The "Stubbornness" Meter (The Switching Mechanism)

If the robot just follows you, you might lead it into a trap. If the robot just forces you to follow it, you might get annoyed and stop cooperating. The paper introduces a clever way to decide who is in charge: The Stubbornness Meter.

The Analogy: Think of the robot as a GPS and you as the driver.
- Phase 1 (Robot Leads): At the start, the robot suggests a route. If you agree, great. But if the robot keeps suggesting routes that feel "wrong" to you (like squeezing through a tiny gap), your internal "annoyance meter" starts to rise.
- Phase 2 (The Tipping Point): The robot tracks this annoyance. As long as you are just slightly annoyed, the robot keeps trying to guide you. But once your annoyance crosses a certain threshold (you get "stubborn"), the robot says, "Okay, I give up. You lead, I'll follow."
- The Result: This prevents the team from getting stuck in a loop of disagreement. The robot leads when it's efficient, but hands over the reins the moment you get too frustrated, ensuring the team keeps moving.

3. The "Pre-emptive Dance" (Pose Optimization)

Here is the tricky part: Even after the robot decides to let you lead, it still doesn't know exactly which path you will pick next. You might turn left, or you might turn right. If the robot waits until you turn to move its arms, it might be too late, causing jerky, unsafe movements.

The Solution: The robot uses its "cloud of possibilities" to prepare for all likely outcomes at once.

The Analogy: Imagine you are a dancer following a partner who hasn't decided which step to take next. Instead of standing stiffly, you shift your weight slightly so that you are balanced and ready to spin either left or right instantly. You aren't committed to one move, but you are ready for any of them.
How it works: The robot adjusts its joint angles (its "pose") to a "sweet spot" that works well for all the paths you might choose. This way, no matter which way you suddenly decide to go, the robot is already in a comfortable, safe position to follow smoothly, avoiding jerky stops or dangerous stretches.

How They Proved It Worked

The researchers didn't just write code; they tested it:

Human Feedback: They showed 20 people videos of different paths and asked them what they would choose. They used this data to tune their "cloud of possibilities" model, proving it could accurately predict how humans distribute their choices.
Simulations: They ran thousands of computer simulations to test the "Stubbornness Meter." They found that teams using this switching mechanism got the job done faster and safer than teams where the robot always tried to lead or always blindly followed.
Control Tests: They showed that the "Pre-emptive Dance" (Pose Optimization) significantly reduced the "jerkiness" and effort required by the robot when following a human, compared to robots that didn't prepare in advance.

The Bottom Line

This paper presents a framework where a robot doesn't just blindly follow a human or blindly force a human to follow it. Instead, the robot:

Acknowledges it doesn't know exactly what you want (Uncertainty).
Waits to see if you get annoyed enough to take charge (Stubbornness).
Physically prepares its body to handle whatever you decide to do next (Pose Optimization).

The result is a team that moves together more smoothly, safely, and efficiently, even without saying a word.

Technical Summary: Mutual Adaptation in Human-Robot Co-Transportation with Human Preference Uncertainty

Problem Statement

The paper addresses the challenges of human-robot co-transportation, specifically focusing on scenarios where a human and a robot jointly navigate an object through obstacle openings to a target. The core problem lies in the uncertainty of human preference parameters and the difficulty of balancing adaptation strategies.

Traditional approaches often rely on fixed parameters to model human behavior, which fails to account for individual variations and fluctuations in risk perception and distance sensitivity. Furthermore, existing collaboration models frequently focus on static, one-time interactions or single-direction adaptation (either robot adapting to human or human adapting to robot), rather than dynamic, mutual adaptation in evolving environments. The authors identify three specific gaps:

Lack of models that quantify human decision-making uncertainties arising from subjective preference parameters.
Absence of mechanisms for dynamic role shifting (who leads the team) based on human resistance to the robot's plan.
Insufficient low-level control strategies to mitigate the impact of uncertain human choices when the human is leading.

Methodology

The authors propose a unified framework comprising three interconnected modules, each addressing a specific hypothesis regarding human preference uncertainty, mutual adaptation, and low-level control.

1. Human Preference Model with Parameter Uncertainty (H1)

Instead of assuming fixed human parameters, the authors model human decision-making as a probability distribution.

Subjective Cost: They define a subjective cost function $\hat{J}_o(\alpha, \gamma)$ based on Prelec's probability weighting function (for risk perception, parameter $\gamma$ ) and Stevens' power law (for distance perception, parameter $\alpha$ ).
Uncertainty Modeling: Parameters $\alpha$ and $\gamma$ are treated as independent variables following uniform distributions ( $\alpha \sim U(0.55, 0.95)$ , $\gamma \sim U(0.5, 1.1)$ ).
Choice Distribution: By integrating over the parameter space, the model computes the probability $P_o$ that a human will choose a specific path option $o$ . This converts parameter uncertainty into a choice distribution, allowing the robot to anticipate a range of human preferences rather than a single deterministic path.

2. Cooperative Path Planning via Time-Varying Stubbornness (H2)

To manage the transition between robot-led and human-led modes, the framework introduces a time-varying stubbornness measure $\phi(k)$ .

Accumulation of Discomfort: The measure grows exponentially based on the accumulation of "discomfort" $g(k)$ , which occurs when the team's movement deviates from the human's probabilistic preferences.
Mode Transition:
- Robot-Leading Mode ( $\phi(k) \le 1$ ): The robot selects a path that balances its objective cost with the approximated human subjective cost. The weight of the human's cost increases as $\phi(k)$ approaches 1.
- Human-Leading Mode ( $\phi(k) > 1$ ): Once the human's stubbornness exceeds a threshold (1), the robot yields control and follows the human.
Dynamic Adaptation: This mechanism allows the team to start with robot guidance but transition to human leadership if the human consistently resists the robot's plan, thereby achieving mutual adaptation.

3. Low-Level Control via Pose Optimization (H3)

When the robot is in human-leading mode, it must prepare for multiple potential human choices without knowing the exact path the human will take.

Redundancy Resolution: The robot utilizes the redundancy of its manipulator arm.
Probabilistic Pose Optimization: The system solves an optimization problem to select a joint configuration $q^*$ that minimizes the probability-weighted sum of Model Predictive Control (MPC) costs across all possible path options.
Objective: This proactive adjustment ensures the robot maintains favorable kinematic configurations to handle unexpected switches between path options, reducing control effort and tracking errors.

Key Contributions

Probabilistic Human Preference Model: A mathematical framework that captures uncertainties in human risk and distance perception parameters, generating a probability distribution of human choices rather than relying on fixed values.
Mutual Adaptation via Stubbornness: A coordination model using a time-varying stubbornness measure to dynamically transition between robot-leading and human-leading modes, enabling the team to adapt to human resistance over time.
Pose Optimization for Uncertainty: A low-level control strategy that uses the probability distribution of human choices to optimize the robot's joint configuration, mitigating the impact of uncertain human decisions during human-led transport.

Experimental Results

The framework was validated through human participant studies and simulations:

Validation of H1 (Preference Model): Experiments with 20 participants in four different obstacle environments showed that the proposed probabilistic model accurately captured the ranking of path options chosen by humans. The model successfully predicted the likelihood of humans choosing safer but longer paths versus riskier, shorter ones.
Validation of H2 (Mutual Adaptation): Simulations comparing the proposed mode-switching mechanism against a fixed-mode baseline (where the robot always leads or always follows) demonstrated that the adaptive approach consistently achieved lower overall objective costs. The results indicated that the robot could successfully guide the human toward more efficient paths before the human became sufficiently stubborn to force a mode switch.
Validation of H3 (Pose Optimization): Simulation results across various parameter settings showed that enabling pose optimization in human-led mode significantly reduced the total robot control cost (tracking error and control effort) compared to a baseline without pose optimization. This confirmed that preparing for multiple potential human choices improves low-level control performance.

Significance and Claims

The paper claims that its primary significance lies in providing a unified approach that integrates human modeling with uncertain parameters, mutual adaptation strategies, and low-level control optimization.

Addressing Uncertainty: The work moves beyond deterministic human modeling, acknowledging that human parameters vary and fluctuate, and provides a mechanism to handle this via probability distributions.
Balancing Adaptation: It offers a solution to the trade-off between robot efficiency and human preference, allowing the robot to lead when beneficial but yielding when human resistance (stubbornness) indicates a need for human leadership.
Proactive Control: By linking high-level human intent uncertainty to low-level joint configuration, the system proactively mitigates the risks of unexpected human behavior, ensuring safer and smoother co-transportation.

The authors note limitations, specifically that their human feedback was collected via static images and videos rather than physical interaction, and future work will aim to extend the framework to physical co-transportation studies and multi-robot scenarios.