💻 computer science

FILIC: Dual-Loop Force-Guided Imitation Learning with Impedance Torque Control for Contact-Rich Manipulation Tasks

The paper proposes FILIC, a dual-loop framework combining Transformer-based imitation learning with impedance torque control and a cost-effective end-effector force estimator, to enable precise, compliant, and sensor-free force regulation in contact-rich robotic manipulation tasks.

Original authors: Haizhou Ge, Yufei Jia, Zheng Li, Yue Li, Zhixing Chen, Lu Shi, Lei Han, Ruqi Huang, Guyue Zhou

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Haizhou Ge, Yufei Jia, Zheng Li, Yue Li, Zhixing Chen, Lu Shi, Lei Han, Ruqi Huang, Guyue Zhou

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 teaching a robot to do delicate tasks like screwing in a bottle cap, plugging in a charger, or tightening a clamp. The old way of teaching robots was like teaching a blindfolded person to walk through a crowded room: you told them exactly where to move their feet (position), but you didn't tell them what they were bumping into. If they hit a wall, they just kept pushing harder, often breaking things or damaging themselves.

This paper introduces a new system called FILIC (Force-guided Imitation Learning with Impedance Control). Think of it as giving the robot "super-senses" and "soft hands" without needing expensive, fragile sensors.

Here is how it works, broken down into simple parts:

1. The Problem: Robots are "Blind" to Touch

Most robots learn by watching humans (Imitation Learning). But they usually only learn where to move, not how hard to push.

The Analogy: Imagine trying to thread a needle while wearing thick boxing gloves. You can see the needle, but you can't feel the thread. If you push too hard, you bend the needle; too soft, and you miss.
The Issue: To fix this, engineers usually put expensive "force sensors" on the robot's wrist. But these sensors are costly, break easily, and aren't available on most cheap robots.

2. The Solution: A "Digital Twin" Brain

The authors solved the cost problem by creating a Digital Twin.

The Analogy: Imagine the robot has a perfect, invisible "ghost" version of itself running inside a computer. This ghost knows exactly how heavy the robot's arm is and how gravity pulls on it.
How it works: The real robot moves, and the computer calculates what the robot should feel based on its internal motors. By comparing what the robot actually feels (from its motors) with what the ghost says it should feel, the system can mathematically figure out the "extra" force coming from the outside world (like the resistance of a screw).
The Result: The robot now "knows" it is touching something, even though it has no physical touch sensor on its wrist. It's like guessing you are holding a heavy box by feeling how much your muscles are straining, rather than weighing the box on a scale.

3. The Two-Loop System: The Brain and The Reflex

FILIC uses a "Dual-Loop" architecture, which is like having a conscious brain and a reflexive nervous system working together.

The Outer Loop (The Brain - 25 times a second):
This is the "Imitation Learning" part. It looks at camera images and the estimated "touch" data (from the Digital Twin). It decides the next big move, like "move the hand slightly to the left." It's smart and learns from human demonstrations.
The Inner Loop (The Reflex - 2,000 times a second):
This is the "Impedance Control" part. It takes the brain's command and executes it with "soft hands." Instead of saying "Go to position X," it says "Push gently until you feel resistance, then stop."
- The Analogy: Think of a spring. If you push a spring, it gives way. If you push a wall, it pushes back. The inner loop makes the robot act like a spring. If it hits a wall, it doesn't crash; it gently yields, adjusts, and tries again. This happens so fast (2,000 times a second) that the robot feels incredibly smooth and safe.

4. Teaching with "Vibration"

To teach the robot, the researchers used a special remote control with vibration feedback.

The Analogy: When a human operator was teaching the robot, the remote control vibrated harder when the robot touched something firmly. It was like a "haptic" nudge saying, "Whoa, you're pushing too hard!"
The Result: The human naturally learned to be gentle and precise. The robot recorded these gentle movements, learning not just where to go, but how to feel its way through the task.

5. The Results: Smarter and Safer

The team tested this on four tricky tasks: plugging in an adapter, pressing an emergency button, tightening a C-clamp, and screwing on a bottle cap.

The "Blind" Robot (Position only): Failed often. It would miss the hole or push too hard and break the object.
The "Muscle-Feeling" Robot (Joint Torque): Better, but still clumsy. It could feel the strain in its muscles but didn't understand exactly where the force was coming from.
The FILIC Robot (Estimated Force + Soft Hands): The Winner.
- In the simulation, it succeeded 90% of the time.
- In the real world, it succeeded 80% or more on all tasks.
- It was significantly better than the other methods because it could "feel" the contact clearly and react instantly with its soft, spring-like reflexes.

Summary

FILIC is a way to teach robots to be gentle and precise without buying expensive sensors. It uses a computer "ghost" to guess what the robot is touching and a super-fast "reflex" system to make the robot move like a spring rather than a hammer. This allows robots to perform delicate, contact-heavy tasks safely and successfully, even on cheap hardware.

Technical Summary: FILIC – Dual-Loop Force-Guided Imitation Learning with Impedance Torque Control

Problem Statement

Contact-rich manipulation tasks, such as insertion, assembly, and in-hand manipulation, require precise force regulation to interact safely and effectively with physical environments. While Imitation Learning (IL) has become a dominant paradigm for acquiring such skills, conventional approaches remain largely position-centric. These methods typically output joint positions or 6-DOF poses, which often induce excessive internal forces during contact, risking damage to the robot and the environment.

Although end-effector force feedback is critical for compliant behavior, relying on wrist-mounted Force/Torque (F/T) sensors limits scalability due to high costs and hardware complexity. Furthermore, existing attempts to incorporate force into IL often treat it merely as an auxiliary signal for position control rather than leveraging it for compliant execution. There remains a significant gap in deploying stable, compliant, torque-level control systems in low-cost, sensorless settings.

Methodology: The FILIC Framework

To address these challenges, the authors propose FILIC (Force-guided Imitation Learning with Impedance torque Control), a dual-loop architecture that integrates a high-level Transformer-based IL policy with a low-level impedance controller.

1. Dual-Loop Control Architecture

The system operates on two distinct time scales to balance perception, planning, and execution:

Outer Loop (25 Hz): A Transformer-based IL policy processes multimodal sensory inputs, including visual images (encoded via ResNet), end-effector pose, and estimated external forces. The policy predicts a sequence of target 6-DOF end-effector poses. This loop utilizes an Action Chunking with Transformers (ACT) approach to enhance temporal coherence and robustness.
Inner Loop (2 kHz): The predicted target poses are converted into joint angles via Inverse Kinematics (IK). A joint-space impedance controller then executes torque commands at a high frequency. This loop employs a simplified spring-damper model (omitting the inertial term to avoid acceleration noise) to ensure stable, compliant torque-level execution. A gravity compensation module runs at 250 Hz to support the dynamics estimation.

2. Sensorless End-Effector Force Estimation

A core innovation of FILIC is its ability to estimate end-effector wrenches without dedicated F/T sensors.

Jacobian-Based Inversion: The system leverages the analytical relationship between joint torques and end-effector forces: $\tau_{ext} = J(q)^T F_{ext}$ . By inverting the Jacobian transpose, the system estimates external forces from joint torque measurements.
Synchronous Digital Twin: To handle real-world factors like sensor noise, friction, and mechanical compliance, the authors employ a high-fidelity digital twin (simulated in MuJoCo). This twin is synchronized with the physical robot's joint positions and velocities to compute expected joint torques ( $\tau_{inv}$ ). The external force is then estimated by comparing observed torques ( $\tau_{obs}$ ) with the model-predicted torques:
$F_{ext} = J(q)^T + (\tau_{obs} - \tau_{inv})$
This approach eliminates the need for expensive hardware while providing intuitive force representations for the IL policy.

3. Haptic-Feedback Data Collection

To ensure high-quality demonstration data that captures meaningful force variations, the authors utilize a low-cost handheld teleoperation interface equipped with vibrotactile haptic feedback. The estimated end-effector forces are mapped to vibration frequencies, allowing human operators to intuitively perceive contact intensity and adjust their actions accordingly.

Key Contributions

The paper outlines three primary contributions:

Sensorless Force Estimation: A low-cost method for estimating end-effector wrenches using a synchronous digital twin and Jacobian-based inversion, with an open-source implementation.
Dual-Loop Framework: A novel architecture combining an outer-loop Transformer-based imitation policy (fusing visual and force inputs) with an inner-loop impedance torque controller, enabling compliant manipulation without wrist F/T sensors.
Performance Validation: Empirical evidence demonstrating that replacing raw joint torque with estimated end-effector force significantly improves policy performance in contact-rich tasks.

Experimental Results

The authors evaluated FILIC across simulation and real-world scenarios involving four contact-rich tasks: plug adapter insertion, emergency stop button operation, C-clamp fixation, and bottle cap tightening.

Force Estimation Accuracy: Static load tests on a 6-DOF manipulator confirmed the estimator's ability to track gravitational loads with low mean error and standard deviation.
Policy Performance (Simulation):
- Position-only baseline: 68.0% success rate.
- Position + Joint Torque: 80.0% success rate.
- Position + Estimated EE Force (FILIC): 90.0% success rate.
Policy Performance (Real-World):
- FILIC consistently outperformed both position-only and position-plus-joint-torque baselines across all four tasks.
- Improvements over position-only ranged from +33.3 pp (Plug) to +53.4 pp (Cap).
- Crucially, Estimated EE Force consistently outperformed raw Joint Torque inputs (e.g., +16.7 pp improvement in the Plug task), indicating that the estimated wrench provides a more task-aligned contact signal.

Significance and Claims

The paper claims that FILIC successfully bridges the gap between modern imitation learning and compliant physical interaction without relying on expensive hardware. By unifying a high-level semantic policy with a low-level torque controller and a sensorless force estimator, the framework achieves safer, more compliant, and adaptable contact-rich manipulation.

The authors emphasize that their approach allows robots to exploit rich physical interaction dynamics in resource-constrained settings. The results suggest that explicit force representation, derived via digital twin-based estimation, is superior to raw joint torque for guiding IL policies in contact scenarios. The work concludes that such a framework enables human-like dexterity in unstructured environments while maintaining scalability and accessibility.