🤖 AI

Robots Influencing Humans to Reveal their Goals during Collaboration and Competition

This paper proposes a unified strategy that accelerates human goal inference in collaborative and competitive settings by guiding humans toward Critical Decision Points where divergent strategies maximize information gain, demonstrating superior accuracy and speed over baselines in both simulation and real-world robot experiments.

Original authors: Debasmita Ghose, Oz Gitelson, Michal Lewkowicz, Jake Brawer, Marynel Vazquez, Brian Scassellati

Published 2026-09-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Debasmita Ghose, Oz Gitelson, Michal Lewkowicz, Jake Brawer, Marynel Vazquez, Brian Scassellati

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 a world where robots and people work side by side, not as machines following rigid commands, but as partners trying to understand each other's intentions. For a robot to be a true partner, it must guess what a human is trying to achieve before that person finishes the task. This is difficult because people often start many different tasks in the same way. A person might pick up a spoon to stir soup, mix a salad, or scoop ice cream. To a robot watching from the outside, these early actions look identical, leaving the machine uncertain about the ultimate goal. If the robot guesses wrong, it might hand over the wrong ingredient or block a path, wasting time and frustrating the human. The challenge, then, is not just to watch and wait for the answer to appear, but to find a way to gently guide the human toward a moment where their true intention becomes impossible to miss.

Researchers at Yale University, the Massachusetts Institute of Technology, and the University of Colorado have developed a new way for robots to solve this puzzle. They propose that instead of passively waiting for a human to reveal their goal, a robot should actively steer the interaction toward specific moments they call "Critical Decision Points." These are situations where the path forward splits, and the choices a human makes clearly separate one goal from another. For example, if a person is making either a smoothie or a stew, the early steps of gathering water and vegetables are shared by both. However, the moment the person decides to put those ingredients into a blender or a pot is a critical decision point. The researchers found that if a robot can subtly influence the human to reach these split-second moments sooner, the robot can figure out the goal much faster and with greater accuracy than if it simply watched.

To test this idea, the team created two very different scenarios: a cooperative cooking task and a competitive game of hide-and-seek. In the cooking experiment, a human and a robot worked together in a kitchen to prepare one of thirty possible meals, ranging from oatmeal to pasta. The researchers programmed the robot with a planning system that looked ahead at possible future actions. The robot would choose moves that kept the task moving forward while also setting up the human to make a choice that would clarify their intent. For instance, if the robot needed to know whether the human wanted a parfait or a smoothie, it might first gather a general item like a spoon or a bowl that could be used for either. This kept the robot helpful without committing to a specific recipe. By doing so, it encouraged the human to eventually pick a specific ingredient, like yogurt or kale, which would immediately reveal the intended dish.

The results of these cooking simulations and real-world tests with a physical robot were striking. The new method allowed the robot to identify the correct meal significantly earlier than other strategies. In many cases, the robot could guess the goal correctly within the first third of the interaction, whereas other methods often struggled until the task was nearly halfway done. When the robot guessed early, it made fewer mistakes, such as adding the wrong ingredients or taking unnecessary steps. In contrast, other approaches that relied on passive observation or tried to maximize information gathering without considering the task often led the robot to take wrong turns, forcing the human to correct them later. The study suggests that by balancing the need to finish the job with the need to learn the human's plan, the robot becomes a more efficient and reliable partner.

The researchers also tested their approach in a competitive setting, where a robot played hide-and-seek against a human-controlled robot. In this game, the human tried to hide, and the robot seeker had to find them. Because the human could hide behind walls, the robot could not always see them, making it hard to know where they were going. Here, the "Critical Decision Points" were specific locations on the map where the human's possible hiding strategies would force them to take different paths. The robot used its planning system to move in a way that encouraged the human to head toward these decision points. By doing so, the robot could observe the human's next move and deduce their hiding strategy much faster. In simulations and real-world trials, the robot using this method caught the hider in fewer steps than robots that simply chased the human or explored the map randomly.

This work highlights a shift in how robots interact with people. Instead of treating humans as unpredictable variables to be observed, the robot becomes an active participant that shapes the environment to make human intentions clearer. The researchers found that this approach works whether the human is a helpful partner or a competitive opponent, and whether the robot can see everything or only parts of the scene. While the current experiments were conducted in controlled environments with specific rules, the findings suggest a powerful principle: robots can learn faster and work better by guiding interactions toward the moments where human choices matter most. The study does not claim to have solved every problem of human-robot interaction, but it offers a concrete method for reducing uncertainty and building more natural, effective partnerships between people and machines.

Technical Summary: Robots Influencing Humans to Reveal their Goals during Collaboration and Competition

Problem Statement

Robots operating in shared environments with humans must infer high-level human intentions or goals to act efficiently. However, goal inference is challenging due to partial observability, noisy sensing, and, most critically, the fact that many distinct human goals share overlapping initial action sequences. Standard passive observation methods, such as Bayesian belief updates, often converge too slowly under these conditions, leaving the robot uncertain for a significant portion of the interaction. Furthermore, reliance on verbal communication is often impractical due to physical constraints, task complexity, or the burden it places on the human. The paper addresses the need for a unified strategy that allows robots to actively influence human behavior to reveal their goals earlier in the interaction, applicable to both collaborative and competitive scenarios.

Methodology

The authors propose a unified framework centered on Critical Decision Points (CDPs) and Receding-Horizon Planning (RHP).

1. Critical Decision Points (CDPs)

A CDP is defined as a state in the shared environment where competing human strategies (policies associated with different candidate goals) prescribe significantly different immediate actions.

Formalization: The authors quantify the "informativeness" of a state $s$ using a goal-conditioned policy divergence measure $M(s, H)$ . This measure calculates the total pairwise disagreement between the action distributions induced by policies for different candidate goals $g \in G$ .
Identification: A state is identified as a CDP if the divergence $M(s, H)$ exceeds a threshold $\theta$ . Observing the human's action at a CDP provides maximal information to distinguish between competing goals.

2. Receding-Horizon Planning (RHP) Strategy

To actively drive the interaction toward CDPs, the robot employs an RHP approach. Instead of passively waiting for the human to reach a CDP, the robot plans a sequence of actions to steer the human toward these states while balancing task progress.

Tree Expansion: The robot expands a search tree of depth $D_H$ , alternating between robot action branches and human response branches. Human responses are sampled from the policy bank $\bar{\pi}_g$ corresponding to the robot's current belief about the human's goal.
Cost Function: The planner evaluates leaf nodes using a weighted cost function $J$ $J$ :
$J = \lambda_1 C_{task} + \lambda_2 M_{informativeness}$
- Task Cost ( $C_{task}$ ): Ensures the robot's actions support the task (e.g., selecting general actions in collaboration or minimizing distance in competition).
- Informativeness Cost ( $M_{informativeness}$ ): Measures the divergence of goal-conditioned policies at the leaf state. In the collaborative setting, this is calculated as the dissimilarity between the belief over goals given the history and the belief given the future sequence. In the competitive setting, it is the distance to the nearest pre-computed CDP.
Execution: The robot executes the first action of the path that minimizes $J$ , observes the human's response, updates its belief via Bayes' rule, and replans.

3. Application Scenarios

The framework is validated in two distinct settings:

Scenario 1: Collaborative Cooking (Fully Observable): A human and robot take turns preparing one of 30 recipes. Many recipes share initial steps (e.g., gathering water, mixing). The robot uses RHP to select general actions that keep options open while pushing the human toward specific ingredients (CDPs) that disambiguate the recipe.
Scenario 2: Competitive Hide-and-Seek (Partially Observable): A seeker robot attempts to catch a human-controlled hider. The hider's strategy is hidden. The robot pre-computes CDPs in the grid world and uses RHP to maneuver the hider into these states (often by triggering evasive maneuvers) to observe the hider's underlying policy.

Key Contributions

Identification of CDPs: The paper demonstrates that human intentions can be revealed significantly earlier if the robot actively drives the interaction toward a small set of highly informative states (CDPs), defined as states where competing strategies prescribe different actions.
Unified Framework: The authors provide a formal, goal-conditioned definition of CDPs and an accompanying RHP planner that balances task progress and information gain. This is the first unified goal-prediction approach shown to operate effectively across both cooperative (fully observable) and competitive (partially observable) interactions.
Empirical Validation: The approach is evaluated in both simulation and on real robots (UR-5e for cooking, Turtlebot3 for hide-and-seek), showing superior performance over baselines in terms of inference speed and accuracy.

Results

The proposed method was compared against Bayesian Delegation (a passive belief-updating baseline) and Information Gain Maximization (an active baseline that maximizes immediate information gain without considering task alignment).

Collaborative Cooking:
- Goal Inference Speed: The proposed method identified the human's goal significantly earlier than baselines. The "first correct guess" occurred in the first third of the interaction, and the "last incorrect guess" was resolved around the midpoint, whereas baselines remained uncertain longer.
- Efficiency: The method resulted in the fewest "extra steps" (0.70 steps on average) compared to Bayesian Delegation (11.0 steps) and Information Gain Maximization (7.02 steps).
- Ablation: Removing the temporal history (Vanilla CDP) or the task-progress term led to suboptimal performance, confirming the necessity of both history-awareness and task alignment.
Competitive Hide-and-Seek:
- Capture Time: In simulation and real-world trials, the seeker using the CDP-RHP strategy caught the hider faster than a random explorer or a seeker that simply minimized distance to the hider.
- Strategy Inference: By steering the hider toward CDPs, the seeker could observe the hider's underlying policy more quickly, allowing for faster interception.

Significance and Claims

The paper claims that the core insight—actively steering interactions toward states where human strategies diverge—provides a robust mechanism for fast goal inference.

Overcoming Overlap: The method specifically addresses the challenge of overlapping action sequences, which often confuses passive observers. By selecting actions that are "general" enough to be safe but "specific" enough to force a decision, the robot accelerates disambiguation.
Versatility: The framework is significant because it does not rely on a fixed interaction mode (collaborative vs. competitive) or observability level. It adapts by changing how the cost function is weighted and how CDPs are computed (online vs. pre-computed).
Active vs. Passive: The authors argue that fast goal inference depends not just on how the robot updates beliefs, but on whether the robot can guide the interaction to situations where informative evidence becomes available. The CDP framework enables the robot to act in ways that make future human behavior more diagnostic.

The authors remain modest regarding limitations, acknowledging that the current approach assumes deterministic environments, rational human actors, turn-based interactions, and the availability of explicit policy banks. They suggest future work should address goal changes, continuous state spaces, and the integration of language-based interactions.