PaperlayerView exact revision on arXiv

Source-linked AI summary

Robots Influencing Humans to Reveal their Goals during Collaboration and Competition

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

arXiv:2609.05519v1cs.ROcs.AI

TL;DR

Robots need to infer human goals despite ambiguous, noisy, and partially observable interactions. This paper drives humans toward Critical Decision Points using a Receding-Horizon Planner that balances task progress with information gain, and reports earlier, more accurate inference across collaborative cooking and competitive hide-and-seek tasks.

  • Problem

    Human goals are difficult to infer quickly when observations are sparse or ambiguous, goals share initial action sequences, and environments may be partially observable.

  • Method

    The method identifies Critical Decision Points through goal-conditioned policy divergence and uses Receding-Horizon Planning to steer interactions toward them while balancing task progress and information gain.

  • Results

    The framework yields earlier and more accurate human-goal inference than baselines across collaborative cooking and competitive hide-and-seek tasks.

  • Takeaways & Limitations

    The results support using active interaction with shared-environment structure for goal inference across cooperative and competitive settings, including partial observability.

  • Takeaways & Limitations

    The experiments assume deterministic, turn-based, discrete environments with rational humans, fixed tasks, and an explicit policy bank.

Abstract

from arXiv · show

We propose a unified strategy for fast goal inference in human-robot interaction. The core idea is to drive the human toward Critical Decision Points (CDPs)-states where competing human strategies prescribe different next actions and thus maximally reveal the goal. We formalise CDPs using a goal-conditioned policy divergence measure and incorporate them into a Receding-Horizon Planner that explores future action sequences while optimizing a cost function balancing task progress and information gain. We evaluate this approach in both a collaborative, fully observable cooking task and a competitive, partially observable hide-and-seek game, each in simulation and on real robots. In both scenarios, our method infers human goals more accurately and earlier than baseline strategies.

1 Introduction

The paper proposes actively steering human-robot interactions toward Critical Decision Points (CDPs), where competing goals imply different next actions. A Receding-Horizon Planner balances task progress and information gain, and the framework is evaluated across collaborative cooking and competitive hide-and-seek settings.

  • Motivation: Goal inference is difficult when observations are sparse, noisy, partially observable, adversarial, or shared across multiple human goals.Passive Bayesian approaches can converge too slowly under these conditions.
  • Core idea: Critical Decision Points are states where competing human strategies prescribe different immediate actions, making subsequent behavior highly revealing of the human’s goal.The paper identifies CDPs as especially informative states for early goal inference.
  • Method: The Receding-Horizon Planner explores possible future human-robot actions and selects actions that maximize information gain while balancing task performance.The planner is designed to guide the interaction toward CDPs rather than rely solely on passive observation.
  • Evaluation: The framework is evaluated in collaborative, fully observable cooking and competitive, partially observable hide-and-seek tasks with contrasting action dependencies.The experiments include simulation and physical-robot settings.
  • Results: The method infers human intentions earlier and more accurately than baselines across cooperative and competitive interactions.In hide-and-seek, influencing the hider toward CDPs enables the seeker to catch the hider more quickly than random exploration or distance minimization.

2 Related Work

Prior work infers human goals through Bayesian methods, supervised learning, opponent modeling, and active information gathering. This paper extends active goal inference with Critical Decision Points for online sequential collaboration involving overlapping action sequences.

  • Human goal inference: Goal-inference research includes Bayesian posterior estimation, dynamic probabilistic models, and supervised neural networks mapping observations to intentions.These approaches can operate over single observations or sequential processes.
  • Active inference: Active strategies have used interventions to disambiguate goals or information-gathering actions without necessarily contributing to shared task completion.Bayesian Delegation instead predicts human responses to robot actions for collaboration.
  • This work: The paper introduces CDPs for quickly inferring human goals online when overlapping action sequences support multiple goals.The approach targets sequential human-robot collaboration rather than only passive observation.
  • Competitive interaction: Opponent modeling builds behavioral models from observed actions and has been applied to stochastic games, negotiations, cooperative tasks, and human-robot interaction.Related human-robot approaches use Theory of Mind to reason recursively about human actions and robot behavior.
  • Influencing behavior: Other work influences human behavior by modeling rewards, steering people toward desired states, inferring strategies from repeated interactions, or using theory-of-mind approaches.These lines of research span both collaborative and competitive scenarios.

3 Influencing Humans to Reveal their Goals at Critical Decision Points

The method models candidate human goals with policy banks, maintains a belief over those goals, and identifies Critical Decision Points where goal-conditioned actions diverge. Receding-Horizon Planning steers interaction toward such states by balancing task progress with information gain.

  • Goal and policy representation: The robot represents each candidate goal with a policy bank of possible human strategies and assumes the goal remains fixed during an interaction.Policies map the current state and, when relevant, interaction history, to distributions over human actions.
  • Belief over goals: The robot updates its belief over goals from observed human actions using a Bayes filter that marginalizes over policies within each goal’s policy bank.The marginal likelihood treats valid policies for a goal as having equal prior weight.
  • Critical Decision Points: A Critical Decision Point is a state where policies associated with different candidate goals prescribe sufficiently different immediate actions to make the next human action informative.Differences among policies pursuing the same goal do not suffice unless they separate that goal from other candidates.
  • Critical Decision Points: The disagreement measure compares marginal action distributions across candidate goals, with high disagreement indicating states where observing the human’s next action can distinguish goals.In the illustrative example, distinct actions at s1 produce high disagreement, whereas identical actions at s3 produce zero divergence.
  • Receding-Horizon Planning: The RHP planner expands possible future human-robot action sequences and selects actions using a weighted combination of task cost and informativeness cost.The informativeness cost reuses the divergence measure defining CDPs, while the planner executes the first action on the lowest-cost branch before observing and updating belief.

4 Scenario 1: Influencing the Human to Reveal their Goal in Human-Robot Collaboration

In collaborative cooking, the robot uses RHP to choose supportive actions that preserve ambiguity until the human reaches an informative CDP. It plans over likely future branches, balancing general task-compatible actions against actions expected to change goal belief.

  • RHP tree expansion: The planner expands a turn-based RHP tree from the current interaction history, alternating robot actions with sampled human actions over a finite horizon.Each leaf stores its resulting state, augmented history, and simulated future action sequence for scoring.
  • Approach: The robot can complete shared preparatory work to expose a later human choice that distinguishes competing meals, such as adding kale to a pan or bowl.The cooking example uses the robot’s action to bring the interaction to a CDP before observing the human’s clarifying action.
  • Action selection: The selected robot action is the first action on the lowest-cost RHP branch, after which the robot observes the human and updates its belief.The cost combines task and informativeness terms with empirically determined weights.
  • Action generality: The planner favors general actions compatible with many goals because consuming those actions reduces the human’s remaining generic options and encourages more specific choices.Gathering a spoon or water can support several recipes, whereas pouring an ingredient into a particular container commits to fewer recipes.
  • Belief change: The planner also evaluates future action sequences by how strongly they alter the belief over human goals, using a dissimilarity measure computed from belief distributions.The dissimilarity can be calculated online when the state space is too large for offline evaluation.

4.2 Task and Problem Representation

The collaborative cooking task models human-robot meal preparation as a turn-based planning problem with many recipes sharing preparatory actions. A symbolic PDDL environment supports both simulated and physical experiments across 30 possible recipes.

  • Task motivation: The cooking benchmark includes many recipes whose common preparatory actions make the intended meal difficult to infer without distinguishing actions.Chopping, boiling, and mixing can occur across multiple recipes, creating ambiguity under incomplete information.
  • Recipe set: The recipe set spans meals including pasta, stew, salad, oatmeal, smoothie, and parfait, with ingredient requirements listed in Table 1.The table organizes the candidate recipes and their required ingredients for the collaborative setup.
  • Experimental setup: The experiments cover simulation and real-world collaboration in which humans and robots take turns preparing one of 30 possible recipes.The setup assumes either agent can perform required preparatory steps, with each step taking the same time.
  • Task representation: The environment represents robot and human actions with PDDL, dynamically updating state as either agent performs actions.The action set includes gathering, pouring, mixing, cooking, collecting water, blending, heating, and serving.

4.3 Experimental Setup

The collaborative cooking experiments model human goals with ground-truth action sequences and approximate goal beliefs, while constraining planning through action-probability estimates and fallback behavior.

  • Simulation: The simulation represents human policies as ground-truth action sequences derived from Clique/Chain Hierarchical Task Networks.These networks encode both sequential dependencies and unordered subtasks.
  • Simulation: Goal-belief computation is approximated because unordered recipe steps create prohibitively many possible action sequences.
  • Planning: The RHP tree uses offline n-gram action probabilities to limit expansion when many actions are valid at each timestep.
  • Planning: When an action sequence is absent from the n-gram data, planning falls back to branches supporting the most likely goals.
  • Planning: The cost function sets λ1=1 and λ2=10, empirically prioritizing informative interactions while penalizing actions unhelpful for task completion.
  • Physical setup: The physical setup aligns real-world perception and language grounding with the symbolic interface used in simulation.It uses a UR-5e robot, an overhead Azure Kinect, and natural-language action descriptions mapped to PDDL literals.

4.4 Comparisons

The study compares the proposed approach with Bayesian Delegation and information-gain maximization, which infer goals through belief updates or expected information gain rather than explicitly targeting CDPs.

  • Baselines: The evaluation compares the proposed approach with Bayesian Delegation and Information Gain Maximization.
  • Bayesian Delegation: Bayesian Delegation updates its estimate of the human’s goal from observed actions while selecting robot actions that move collaboration forward.
  • Information Gain Maximization: Information Gain Maximization plans over an RHP tree of depth 4 and optimizes expected changes in goal probabilities between root and leaves.
  • Comparison: Unlike CDP-based informativeness, the information-gain baseline varies its measure with the robot’s current uncertainty and may overexplore unlikely goal-revealing states.

4.5 Metrics

The collaborative cooking evaluation measures how early and reliably the robot infers goals, along with the extra work required to complete each interaction.

  • Metrics: The evaluation uses four metrics: first correct guess, last incorrect guess, number of extra steps, and percentage of correct guesses.
  • Metrics: First correct guess records the first timestep when the robot correctly identifies the human’s goal.
  • Metrics: Last incorrect guess records the final timestep when the robot incorrectly identifies the human’s goal.
  • Metrics: Number of extra steps compares joint interaction length with the ground-truth steps from the corresponding HTN.
  • Metrics: The study reports top-1 and top-3 accuracy for first correct guess, last incorrect guess, and percentage of correct guesses.

4.6 Simulation Results

In simulated collaborative cooking, the proposed approach identifies and stabilizes the human’s goal earlier than baselines while requiring few additional interaction steps.

  • Simulation results: 18.7 steps is the average ground-truth interaction length across 30 recipes and three evaluation trials.
  • Goal inference: The proposed approach typically makes its first correct guess in the first third of the sequence, earlier than all baseline approaches.The authors caution that this metric can be noisy because early correct predictions may occur by chance.
  • Goal inference: The proposed approach generally reaches its last incorrect guess around the sequence midpoint, earlier than all baselines.
  • Goal inference: Resolving uncertainty earlier produces a higher percentage of correct guesses for the proposed approach.
  • Task efficiency: 0.70 extra steps is the proposed approach’s average overhead in completing the collaboration.The ground-truth baseline is 18.7 steps, while incorrect robot actions could extend the interaction to nearly twice that length.

4.7 Ablation Study

The ablations show that both generality and similarity are needed for strong performance, while temporal modeling improves CDP-based goal identification by using interaction history.

  • Cost Function Ablation: The full cost function outperforms either the generality-only or similarity-only condition across all metrics, indicating both terms are necessary.The weighted cost function balances task progress and information gain through λ1 and λ2.
  • Cost Function Ablation: Generality alone improves goal prediction but causes overly general actions, task deviation, and unnecessary steps.The robot avoids task-specific actions while remaining compatible with many possible goals.
  • Cost Function Ablation: Similarity alone selects actions beneficial to the human but performs poorly at guiding the interaction toward informative CDPs.It takes fewer unnecessary steps but struggles with goal prediction.
  • Temporal Modeling Ablation: The history-agnostic Vanilla CDP baseline evaluates each state independently, whereas the proposed method conditions informativeness on the entire interaction history.The baseline chooses immediate information gain without allowing past observations to affect which states are considered revealing.

4.8 Real Robot Results

In physical-robot cooking case studies, the proposed method used general actions to elicit distinguishing human actions early, becoming certain sooner while avoiding the extra steps and mistakes seen with baselines.

  • Experimental Setup: The real-robot comparison evaluated Bayesian Delegation, information-gain maximization, Vanilla CDP, and the proposed method using top-three recipe prediction.Humans took the same first action to prevent an initially revealing action from favoring any method.
  • Overall Results: The proposed method selected general actions that encouraged distinguishing human actions early, reducing uncertainty and avoiding incorrect robot actions.The robot’s strategy used actions common to many recipes before the human performed more informative actions.
  • Overall Results: Across the baseline approaches, the human performed most task steps, contrasting with the robot’s more active role under the proposed method.
  • Case Study: Kale Tomato Smoothie: In the Kale Tomato Smoothie case, the proposed method became certain by step 10 after one extra initial step, while Bayesian Delegation took ten extra steps and made recipe-inconsistent mistakes.Vanilla CDP and information-gain maximization each took seven extra steps in this case study.
  • Case Study: Nut Chocolate Parfait: In the Nut Chocolate Parfait case, the proposed method began with general actions and prompted early distinctive ingredients, whereas information-gain maximization selected overly specific actions and took eight extra steps.The proposed sequence included retrieving the bowl and glass and influencing the human to add yogurt, nuts, and chocolate chips early.

4.9 Comparative Analysis of Performance of CDPs in the Collaborative Task

The comparative analysis attributes baseline weaknesses to overlapping action sequences, myopic information gain, and omission of action history, while the proposed CDP approach combines generality, similarity, and history-aware planning.

  • Bayesian Delegation: Bayesian Delegation struggles with overlapping action sequences because shared actions provide insufficient discriminatory evidence, delaying and weakening goal inference.Belief remains spread across multiple goals rather than concentrating promptly on the correct one.
  • Information Gain Maximization: Information Gain Maximization is shortsighted when goals share actions because it prioritizes immediate uncertainty reduction without balancing exploration against alignment with the likely goal.It may pursue the wrong goal while selecting actions that provide the most information immediately.
  • Vanilla CDP: Vanilla CDP considers only the current state, producing implausible action combinations and compounding errors when contextual history is crucial for distinguishing goals.The informativeness of a state can depend on when it occurs in the interaction.
  • Proposed Approach: The proposed approach actively influences human actions toward information-revealing states rather than passively observing them.This addresses challenges faced by the baseline methods in collaborative goal inference.
  • Proposed Approach: Generality guides the human toward distinguishing actions while preserving compatibility with the human’s true goal, addressing overlapping action sequences.The cost function increases the likelihood that actions remain on policy even when the robot’s belief is incorrect.
  • Proposed Approach: Similarity encourages actions relevant to the human’s goal instead of actions that merely maximize belief reduction, while history prevents implausible combinations.Together these design choices help the robot steer collaboration toward earlier uncertainty reduction.

5 Scenario 2: Influencing the Human to Reveal their Goal in a Competitive Human Robot Task

In competitive hide-and-seek under partial observability, the robot uses Receding-Horizon Planning to steer the human toward Critical Decision Points and infer the hidden strategy. Simulation and real-robot results show faster capture than baseline exploration or distance-minimizing strategies.

  • 5.2.1 Task: The competitive task requires inferring a human-controlled hider’s strategy from limited or unreliable history and immediate state information.The seeker knows possible strategies but must infer which one the hider follows from partial observations.
  • 5.1 Approach: RHP expands future seeker and hider actions over a planning horizon, evaluating branches with task-related and informativeness costs to steer the hider toward CDPs.The cost is weighted by the robot’s belief over possible human policies, and the selected action minimizes cost over the horizon.
  • 5.1 Approach: The hide-and-seek implementation uses a precomputed CDP set, a policy bank, a discrete grid, and alternating seeker and hider branches in the RHP tree.The experimental setup includes four simulated environments and a physically similar laboratory environment.
  • 5.1 Approach: At a critical decision point, the seeker can observe strategy-revealing behavior after inducing an evasion, then select an interception action based on the updated policy estimate.The illustrated rollout plans through t4 but executes only the first action before replanning.
  • 5.5.1 Simulation: Across all four simulation environments, the CDP method outperformed random exploration and minimizing distance to the hider whenever visible.The baselines either failed to estimate the strategy accurately or repeatedly triggered evasive behavior by entering the hider’s field of view.
  • 5.5.2 Real World: In real-robot trials, the method caught the hider in 27 ± 3 timesteps, while Baselines 1 and 2 required 39 ± 2 and 24 ± 1 timesteps longer, respectively.The demonstration included five trials each for six hider policies.
  • 5.6 Conclusion: The method identifies strategies more quickly by observing humans at states where plausible policies diverge, assuming a reasonable policy bank and an unchanged strategy during each game.The approach is demonstrated in simulation and the real world.

6 Discussion

The framework works across collaborative and competitive tasks with contrasting observability, interaction modes, and environment sizes. Its scope is limited by deterministic, rational, turn-based settings, discrete representations, explicit policy banks, and action-only inference.

  • 6 Discussion: The framework spans fully observable collaborative cooking and partially observable competitive navigation despite differences in state-space size, objectives, and interaction assumptions.The authors position this contrast as evidence of framework flexibility across interaction modes and observability levels.
  • 6 Discussion: The method guides interactions toward informative states rather than waiting for passive observations to reveal goal-distinguishing actions.In cooking, supportive actions remain compatible with multiple meals; in hide-and-seek, environmental structure helps expose strategy differences.
  • 6.3 Limitations and Future Work: The experiments assume deterministic environments, rational humans without mistakes, unchanged tasks, and turn-based interaction without waiting actions.These assumptions may not apply directly to real-world settings.
  • 6.3 Limitations and Future Work: The approach requires discrete task representations and explicit policy banks, which may be unavailable in many real-world scenarios.Future work proposes continuous or hybrid spaces and learned, retrieved, or generated goal-conditioned policies.
  • 6.3 Limitations and Future Work: The paper isolates action-based inference, leaving integration with language-based interaction as an important direction for real-world collaboration.Language could initialize or update beliefs while CDPs identify informative action states.
Loading 2609.05519v1…