I've Changed My Mind: Robots Adapting to Changing Human Goals during Collaboration

TL;DR

The paper tackles the problem of human goals changing mid-task in human-robot collaboration by explicitly modeling goal changes through tracking multiple candidate human action sequences against a policy bank $\mathcal{P}$ and a goal bank $\mathcal{G}$. It introduces Differentiating Actions and Receding Horizon Planning (RHP) to actively influence the human to reveal updated goals while assisting toward the new objective, and it maintains multiple plausible histories to detect switches promptly. The approach is evaluated in a collaborative cooking domain with up to 30 recipes, showing faster and more reliable goal identification and improved collaboration efficiency compared with baseline methods like Recursive Bayesian, CDP, and Information Gain Maximization. The results from both simulation and a physical robot demonstrate practical benefits for robust, real-time goal inference and adaptive assistance in dynamic, real-world tasks.

Abstract

For effective human-robot collaboration, a robot must align its actions with human goals, even as they change mid-task. Prior approaches often assume fixed goals, reducing goal prediction to a one-time inference. However, in real-world scenarios, humans frequently shift goals, making it challenging for robots to adapt without explicit communication. We propose a method for detecting goal changes by tracking multiple candidate action sequences and verifying their plausibility against a policy bank. Upon detecting a change, the robot refines its belief in relevant past actions and constructs Receding Horizon Planning (RHP) trees to actively select actions that assist the human while encouraging Differentiating Actions to reveal their updated goal. We evaluate our approach in a collaborative cooking environment with up to 30 unique recipes and compare it to three comparable human goal prediction algorithms. Our method outperforms all baselines, quickly converging to the correct goal after a switch, reducing task completion time, and improving collaboration efficiency.

Figures (4)

• Figure 1: Robot reasoning if the person has changed their mind about making the salad, after observing that the person has picked chocolate.
• Figure 2: Steps to Select Robot Action after Goal Change Detection 1. Detect Goal Change if the actions taken during the collaboration don't match any sequences in $\mathcal{P}$. 2. Create potential action sequences by concatenating the action at which the goal change was detected with past actions sorted by increasing generality scores. 3. Check if the potential action sequences are plausible. 4. For all plausible action sequences, expand the RHP trees. 5. Compute attractor field cost and select an optimal next robot action.
• Figure 3: Comparison of performance measures for the proposed method, CDP ghose2024planning, Information Gain Maximization sadigh2016information, and Recursive Bayesian recursive_bayesian in a collaborative cooking simulation. Metrics include (a) first correct guess after a goal switch, (b) last incorrect guess, (c) number of mistakes, and (d) percentage of correct guesses. Results are shown for three simulated human types: stubborn (no goal change), optimal (stochastic goal updates), and suboptimal (stochastic updates with mistakes). In the stubborn condition, all four methods are compared; in the other two conditions, our method is directly compared with Recursive Bayesian, while the red and blue lines indicate the empirical performance bounds provided by CDP and Information Gain Maximization, respectively.
• Figure 4: Results of the Case Study with the Physical Robot running Recursive Bayesian recursive_bayesian and Our Proposed Algorithm: A group of columns represent the method. For each group, the first column depicts the agent taking action, the second column depicts the action sequences taken by the respective method, the third column denotes the ground truth human goal, and the fourth column the presence and position of the current ground truth goal recipe in the top-3 most probable recipes from the goal prediction model (left-most being the most probable recipe).