TATIC: Task-Aware Temporal Learning for Human Intent Inference from Physical Corrections in Human-Robot Collaboration

This paper introduces TATIC, a unified framework that combines torque-based force estimation with a task-aware Temporal Convolutional Network to infer both discrete task-level intent and continuous motion parameters from brief physical corrections, enabling robust online adaptation in human-robot collaboration.

The Gist

Imagine you are working on a complex puzzle with a very strong, very fast robot assistant. You both need to take apart an old computer to recycle it. The robot is great at following instructions, but it can't read your mind. Sometimes, it might try to pull a part in a direction that would break something, or it might be moving too fast for your comfort.

In the past, if you wanted to stop the robot, you'd have to grab its arm and hold it there, physically wrestling it into a new path for as long as the problem existed. This is tiring and frustrating.

Enter TATIC. Think of TATIC as the robot's new "superpower" to understand quick, subtle nudges instead of long, exhausting struggles.

Here is how the paper explains this breakthrough, broken down into simple concepts:

1. The Problem: The "Mind-Reading" Gap

Robots today are getting smarter. They use cameras and language models (like the AI in your phone) to understand what you say or what you see. But they are terrible at understanding what you feel or touch.

If you give a robot a quick, gentle push to say, "Hey, don't go there, go around that screw," current robots often ignore it or get confused. They either need you to keep pushing them (which is tiring) or they need you to stop and type a new command (which is slow).

2. The Solution: TATIC (The "Nudge Translator")

The researchers built a system called TATIC. Think of it as a translator that turns a split-second physical touch into a clear, logical command.

Instead of just feeling "pressure," TATIC asks the robot:

• What is the intent? (Are you telling me to stop? Slow down? Go left? Or switch to a different task?)
• What are the details? (How far should I move? How much should I slow down?)

3. How It Works: The "Local Map" Trick

One of the hardest parts of this is that the robot might be facing North, South, East, or West. If you push it "forward" while it's facing North, that's different than pushing it "forward" while it's facing South.

TATIC uses a clever trick called Feature Canonicalization.

• The Analogy: Imagine you are giving directions to a friend. Instead of saying "Go North," you say, "Go straight ahead of you."
• TATIC creates a temporary, invisible "local map" that always aligns with the direction the robot is currently moving. It translates your physical push into this local map. This way, a "push to the left" always means "move left relative to the robot's current path," no matter how the robot is twisted or turned in the room. This makes the robot incredibly smart about adapting to different room layouts.

4. The "Five Magic Words"

When you give the robot a quick nudge, TATIC instantly translates it into one of five specific "magic words" (operators):

1. GUIDE: "Move the path slightly this way." (Like gently steering a car around a pothole).
2. YIELD: "Give me more space." (The robot inflates its safety bubble to avoid hitting things).
3. SLOW: "Take it easy." (The robot slows down its speed).
4. STOP: "Hold on!" (The robot freezes in place, but stays ready to go again).
5. SWITCH: "Change the plan." (The robot realizes you want to pick up a different tool or go to a different part of the task).

5. The "Brain" (Temporal Convolutional Network)

To understand these nudges, TATIC uses a special type of AI brain called a Temporal Convolutional Network (TCN).

• The Analogy: Think of listening to a song. If you only hear one single note, you don't know the melody. You need to hear the sequence of notes to understand the tune.
• Similarly, TATIC doesn't just look at the force of your push at one single millisecond. It looks at the sequence of the push over a tiny fraction of a second. This helps it distinguish between a "Stop" (a hard, sudden push) and a "Guide" (a smooth, flowing push).

6. The Real-World Test: Taking Apart a Computer

The researchers tested this on a real robot arm helping a human take apart a desktop computer.

• Scenario: The robot is about to pull a part, but the human sees a screwdriver that might get crushed.
• Action: The human gives the robot's arm a quick, gentle tap to the side.
• Result: TATIC instantly realizes, "Ah, they want to GUIDE the path to the left to avoid the screwdriver." The robot smoothly adjusts its path in real-time without stopping the whole operation.

Why This Matters

This research bridges the gap between high-level thinking (what the robot is trying to achieve) and low-level feeling (how the human is touching the robot).

It allows humans and robots to work together like a dance partner. You don't need to shout instructions or wrestle the robot. You just give a quick, intuitive nudge, and the robot understands exactly what you mean, making collaboration safer, faster, and much less exhausting for humans.

Technical Summary

Here is a detailed technical summary of the paper "TATIC: Task-Aware Temporal Learning for Human Intent Inference from Physical Corrections in Human-Robot Collaboration."

1. Problem Statement

In Human-Robot Collaboration (HRC), robots must adapt to dynamic task constraints and evolving human intent. While Physical Human-Robot Interaction (pHRI) offers a natural channel for operators to correct robot trajectories via physical contact, current methods face a significant gap:

• Foundation Models (VLA/LLM): Rely heavily on vision and language, often ignoring force feedback, making them inadequate for scenarios with visual occlusions or high-precision manipulation.
• Traditional pHRI: Focuses on low-level trajectory deformation (kinesthetic guidance) but struggles to extract high-level semantic intent (e.g., "stop," "slow down," "switch target") from brief interactions.
• The Challenge: Existing systems require sustained physical guidance, causing operator fatigue, or fail to interpret brief, transient physical corrections as meaningful task-level commands. There is a need for a framework that can jointly infer discrete task-level intent and continuous motion parameters from short-duration physical contacts without external force sensors.

2. Methodology: The TATIC Framework

The authors propose TATIC (Task-Aware Temporal Learning), a unified framework that decodes semantic intent and motion corrections from brief physical interactions using torque-based estimation and a Temporal Convolutional Network (TCN).

A. Torque-Based Contact Estimation

Instead of relying on external force sensors, TATIC estimates contact forces using the robot's internal joint torque measurements:

1. Residual Torque: Calculates external torque ($\hat{\tau}_{ext}$) by subtracting the modeled dynamics (inertia, Coriolis, gravity) from measured joint torques.
2. Contact Detection: Uses a statistical threshold on the weighted residual torque to detect contact onset and duration.
3. Localization & Force Estimation: Identifies the contacted link and estimates the contact location (normalized scalar $s$) and force vector ($f$) by solving a constrained least-squares optimization problem over a short time window.

B. Task-Aligned Feature Canonicalization

To ensure robustness across different workspace layouts and robot poses, the system projects interaction data into a canonical local frame ($L[t]$):

• The frame is aligned with the robot's reference motion direction and the world vertical axis.
• Features are transformed to remove global translation and yaw dependencies while preserving task-relevant geometry (e.g., force direction relative to the path, distance to workspace boundaries).
• This allows the model to generalize to new environments without retraining.

C. Temporal Learning with Causal TCN

The core inference engine is a Causal Temporal Convolutional Network (TCN) that processes a sliding window of feature vectors ($X[t]$).

• Input Features: Includes kinematic data (speed, force direction/magnitude), workspace constraints (clearance to boundaries), alignment scores (force vs. goal direction), and a query context variable (indicating task ambiguity).
• Architecture: Uses residual blocks with causal dilated convolutions to capture temporal dependencies over the interaction window.
• Multi-Task Output: The network jointly predicts:
  • Discrete Intent: A semantic operator class (GUIDE, YIELD, SLOW, STOP, SWITCH).
  • Continuous Parameters: Task-conditioned values such as displacement direction/magnitude (for GUIDE), safety margin (for YIELD), speed scaling (for SLOW), or target selection (for SWITCH).
• Optimization: Uses a multi-objective loss function with homoscedastic uncertainty weighting to balance heterogeneous tasks (classification vs. regression).

D. Intent-Driven Motion Adaptation

Once intent is inferred, the system adapts the robot's motion in real-time:

• GUIDE: Deforms the nominal path with a smooth "bump function" to avoid obstacles or adjust trajectory.
• YIELD: Inflates the safety buffer (Minkowski sum) around the robot to increase clearance.
• SLOW: Scales the reference velocity.
• SWITCH: Selects a new mission goal from candidate targets.
• STOP: Pauses execution while maintaining control, allowing for quick resumption.

3. Key Contributions

1. Unified Framework: TATIC is the first framework to jointly decode discrete semantic intent and continuous motion parameters from brief physical corrections, bridging the gap between high-level planning and low-level control.
2. Task-Aligned Canonicalization: A novel feature projection method that decouples interaction features from global spatial dependencies, ensuring robust generalization across diverse workspace layouts.
3. Sensor-Free Implementation: Utilizes torque-based contact estimation, eliminating the need for external force/torque sensors and reducing operator fatigue by supporting brief, rather than sustained, interactions.
4. Hardware Validation: Successfully demonstrated on a 7-DoF manipulator in a collaborative disassembly task, proving the system's ability to handle real-world occlusions and dynamic intent changes.

4. Experimental Results

Experiments were conducted on a dataset of 500 in-distribution (ID) episodes and 250 out-of-distribution (OOD) episodes involving collaborative desktop disassembly.

• Intent Recognition Performance:
  • Achieved a Macro-F1 score of 0.904 on the test set.
  • SWITCH intent was recognized with the highest accuracy (F1 = 0.951).
  • GUIDE and YIELD were the most challenging to distinguish (F1 $\approx$ 0.86–0.88) due to similar lateral pushing behaviors, but the system still performed robustly.
• Ablation Studies:
  • Feature Importance: Removing workspace and alignment features dropped the Macro-F1 from 0.904 to 0.583, highlighting the necessity of geometric context.
  • Canonicalization: The proposed canonical frame significantly outperformed world-frame baselines in OOD scenarios (F1 0.871 vs. 0.614), proving its effectiveness in handling layout reconfigurations.
  • Temporal Modeling: The causal TCN significantly outperformed non-temporal baselines (MLP, Linear models) in regression tasks, particularly for direction prediction (Cosine Similarity 0.891 vs. 0.641).
• Hardware Validation: The system successfully executed closed-loop adaptation in a collaborative disassembly task, where brief physical taps corrected the robot's path, adjusted speed, or switched targets in response to human needs (e.g., avoiding collisions, gathering tools).

5. Significance

This work represents a significant step forward in making robots more responsive and intuitive collaborators. By enabling robots to understand what a human wants (semantic intent) and how to move (motion parameters) from a split-second touch, TATIC reduces the cognitive and physical burden on human operators. It moves beyond simple trajectory following to intent-aware collaboration, making HRC viable for complex, unstructured tasks like disassembly and assembly where visual data is often insufficient. The framework's ability to generalize across different environments without retraining is particularly valuable for deploying robots in dynamic industrial and healthcare settings.