🤖 machine learning

Enhancing Goal Inference via Correction Timing

This paper investigates how the timing of human corrections serves as a valuable signal for inferring task-relevant factors, demonstrating that leveraging this temporal information improves the identification of motion features prompting intervention and accelerates the inference of human correction goals.

Original authors: Anjiabei Wang, Shuangge Wang, Tesca Fitzgerald

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anjiabei Wang, Shuangge Wang, Tesca Fitzgerald

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 are teaching a robot to play a game of "fetch" with a ball. The robot has a plan to run, grab the ball, and bring it back. But sometimes, the robot misunderstands the rules or runs the wrong way.

In the past, when a human saw the robot making a mistake, they would physically grab the robot's arm, steer it toward the correct path, and let go. The robot would then look at where the human steered it and try to learn from that new path.

This paper asks a simple but powerful question: What if we also pay attention to when the human decided to grab the robot?

The "Stop Sign" Analogy

Think of the robot's movement like a car driving down a highway.

The Robot's Plan: The car is driving on a set route.
The Human Correction: You, the passenger, reach over and turn the steering wheel because you see a pothole or a wrong exit coming up.

Most researchers only looked at where you turned the wheel (the new path). They ignored when you grabbed the wheel.

This paper argues that the timing of your grab is a secret code.

If you grab the wheel immediately when the car starts to drift, it means you have very high standards for how the car should drive.
If you wait until the car is almost off the road before grabbing, it means you are more patient, or perhaps you were waiting to see if the car could fix itself.

The authors call this "Correction Timing." They believe that the split-second decision to intervene tells the robot more about what you want than just the physical movement of your hand.

The Three Big Questions

The researchers tested three main ideas using a robot arm and 120 human volunteers:

1. What makes people grab the robot?
They wanted to know: Is it because the robot is moving too slow? Too fast? Is it moving in a weird, jerky way? Or is it just getting too far from the target?

The Discovery: They found that people don't just react to one thing. It's a mix of factors. If the robot moves in a way that feels "unnatural" or inefficient, people grab it sooner. It's like if a waiter walks toward your table but bumps into chairs; you might call them over earlier than if they just walked a little slowly.

2. Can we guess the goal before the correction is finished?
Imagine you grab the robot's arm and start moving it. Can the robot guess where you want it to go while you are still moving it, or does it have to wait until you let go?

The Discovery: Yes! By combining where you grabbed it with when you grabbed it, the robot can guess your goal much faster. It's like if you start turning a steering wheel to the left; the driver knows you want to turn left long before you finish the turn. The "when" helps the robot predict the "where" faster.

3. Does timing help us learn the rules better?
If the robot knows exactly where you let go of the arm, does knowing when you grabbed it help it understand the rules of the game better?

The Discovery: Not really, at least for simple tasks. If you let go of the robot right next to the target hole, the robot already knows the goal perfectly. The "when" doesn't add much new info. However, for complex tasks where the final position isn't obvious, timing might be a huge help.

The "Traffic Light" Metaphor

Think of the robot's learning process like a traffic light system:

Spatial Info (Where): This is the Green Light. It tells the robot, "Go this way."
Timing Info (When): This is the Yellow Light. It tells the robot, "Hey, I'm getting nervous! I'm about to intervene. You're getting close to a mistake."

The paper shows that if the robot only listens to the Green Light (where you moved it), it learns okay. But if it also listens to the Yellow Light (when you got nervous and grabbed it), it learns faster and smarter. It can predict what you want before you even finish showing it.

The Bottom Line

This research is a game-changer for making robots more intuitive. Instead of just being "dumb" machines that only learn from the final result of a human's help, they can now learn from the human's hesitation and urgency.

Old Way: "You moved my arm here, so I will go there next time."
New Way: "You grabbed my arm early and quickly because I was moving too jerkily. I will slow down and smooth out my path next time, even before you tell me exactly where to go."

By paying attention to the timing of human feedback, robots can become better teammates, anticipating our needs and correcting their mistakes before we even have to fully intervene.

1. Problem Definition

The paper addresses the challenge of Learning from Corrections (LfC) in human-robot interaction (HRI). While prior work primarily treats human corrections as spatial demonstrations (new trajectories) or preferences (comparisons between original and corrected paths), it largely overlooks the timing of the human's decision to intervene.

The authors hypothesize that the moment a human decides to physically correct a robot ( $t_c$ ) contains implicit information about:

Why the correction is happening (which motion features triggered the intervention).
What the human's intended goal is (inferring the task objective before the correction is fully completed).
How to learn more precise task constraints.

The core research questions (RQs) are:

RQ1: What specific features of a robot's motion prompt humans to correct it?
RQ2: Can correction timing, combined with the initial direction of the correction, improve the immediate inference of the human's intended goal?
RQ3: Can timing information refine the learning of precise task constraints (e.g., the final goal location) compared to using spatial cues alone?

2. Methodology

The authors propose a two-stage approach using data from a user study involving a 7-DoF Kinova Gen3 robotic arm performing pick-and-place tasks.

A. Data Collection

Setup: 120 participants supervised a robot inserting shapes into color-coded holes. The robot executed pre-planned trajectories (some intentionally sub-optimal or incorrect).
Data: 7,435 interaction episodes and 3,585 correction trajectories.
Key Variables:
- $t_c$ : The timestep when the human intervenes.
- $c_p$ : The grasp position (where the human touches the robot).
- $c'_p$ : The initial correction velocity (direction of the push).
- $c_l$ : The release position (where the human lets go).
- $\xi$ : The robot's original pre-planned trajectory.

B. Stage 1: Predicting Intervention Timing (Addressing RQ1)

The authors trained a Transformer-based model to predict the probability of a human correcting the robot at any given timestep.

Input Features: A 7-dimensional feature vector extracted from the robot's trajectory, including:
- Expectation Alignment: Cosine similarity to optimal velocity/position.
- Directness Alignment: Velocity alignment toward the goal.
- Dynamics: Velocity consistency (smoothness).
- Legibility: How easily the motion reveals intent.
- Task Progress: Distance to goal.
- Optimality: Boltzmann rationality (efficiency relative to the optimal path).
Output: A Cumulative Distribution Function (CDF) representing the likelihood of a correction occurring by time $t$ .
Ablation Study: The model was tested with individual features removed to determine which motion characteristics most strongly influence human intervention decisions.

C. Stage 2: Goal Inference (Addressing RQ2 & RQ3)

The authors developed three models to infer the goal distribution $P(g|\cdot)$ :

WHERE Model (Spatial Baseline): Infers the goal using only spatial cues ( $c_p, c'_p$ $c_{p}, c_{p}^{'}$ for early inference; $c_l$ $c_{l}$ for late inference).
- Uses a Multi-Layer Perceptron (MLP) to predict the release point ( $c_l$ ) from the grasp/velocity, followed by a Gaussian Mixture Model (GMM) to map release points to goals.
WHEN Model (Temporal Only): Infers the goal using only the timing of the intervention ( $t_c$ ) and the trajectory history ( $\xi$ ).
COMBINED Model: Integrates both spatial and temporal information using a weighted Bayesian posterior:
$P_w(g | t_c, c_p, c'_p, \xi) \propto P(t_c | g, \xi)^\alpha \cdot P(c_l | g)^{1-\alpha}$
- For RQ2 (Early Inference): Uses $c_p$ and $c'_p$ (onset cues).
- For RQ3 (Late Inference): Uses $c_l$ (release position).

3. Key Results

A. Timing Prediction Accuracy

The multi-feature Transformer model significantly outperformed a single-feature Boltzmann baseline (which only used optimality) in predicting when corrections occur.
Feature Importance:
- Optimality and Direct Velocity Alignment were critical for mid-to-late trajectory corrections.
- Velocity Consistency was important only for very late corrections (near task completion).
- Surprising Finding: Removing the "Distance to Goal" feature actually improved prediction accuracy at 70-90% completion, suggesting humans rely more on motion cues (alignment/efficiency) than absolute distance when deciding to intervene.

B. Goal Inference Performance

Early Inference (RQ2): When using onset cues ( $c_p, c'_p$ $c_{p}, c_{p}^{'}$ ), the COMBINED model achieved significantly lower Kullback-Leibler Divergence (KLD) (i.e., higher accuracy) than the WHERE or WHEN models alone for corrections occurring at 70%, 80%, and 90% of the trajectory.
- Significance: Timing helps resolve ambiguity in the initial correction direction before the human finishes the correction.
Late Inference (RQ3): When using the release position ( $c_l$ $c_{l}$ ), the COMBINED model did not outperform the WHERE model.
- Reason: In this specific task, the release position was already a near-perfect indicator of the goal. Timing added no new information once the spatial cue was complete.

4. Key Contributions

Feature Analysis: Demonstrated that human intervention decisions are non-Markovian (dependent on trajectory history) and driven by a complex interplay of motion features (efficiency, legibility, alignment) rather than just optimality.
Early Goal Inference: Proved that incorporating correction timing allows robots to infer human intent earlier in the interaction, specifically when only the initial grasp and direction are known.
Limitation Identification: Showed that timing is most valuable when spatial information is incomplete; it offers diminishing returns when the correction is fully completed and the goal is spatially obvious.

5. Significance and Implications

Proactive Assistance: By understanding when and why a human is about to correct a robot, systems can anticipate intent and adjust behavior in real-time, rather than waiting for the correction to finish.
Efficient Learning: The work suggests that robots can learn task constraints more efficiently by analyzing the decision to intervene as a signal, not just the resulting trajectory.
Design Guidelines: Robot motion planners should prioritize features that align with human expectations (legibility, directness) to reduce unnecessary interventions or to make necessary interventions more informative.
Future Directions: The authors note that while timing didn't help in this simple pick-and-place task, it may be crucial for complex tasks where the relationship between correction and goal is less direct. Future work aims to integrate these predictive models into real-time online control loops.