[IROS 2025 Prediction] TER-DAgger: Solving Precision Insertion via Force-Aware Residual Learning

TL;DR

High-precision assembly (like USB or pin insertion) is a "boss fight" for robotics because of covariate shift—small visual errors lead to high-force collisions. TER-DAgger solves this by:

1. Predicting Forces: Using force discrepancy as a "failure alarm."
2. Trajectory Editing: Smoothing human corrections into the robot's existing plan.
3. Impedance Control: Ensuring the robot doesn't break things during contact. The result? A 37% lead over standard fine-tuning and a reliable way for one human to supervise multiple robots.

The Problem: Perception is Not Enough

Current Imitation Learning (IL) models like ACT or Diffusion Policy are great at "looking" at tasks, but they lack "feeling." In contact-rich tasks, a 1mm error doesn't just mean a missed target; it means a jammed component. Existing Human-in-the-Loop (HITL) methods like HG-DAgger require a human to watch the robot like a hawk, which doesn't scale. Furthermore, when a human does take over, the sudden jump in the control signal creates a new "distribution shift" that confuses the model.

Methodology: The Force-Aware Architecture

TER-DAgger uses a two-stage approach. First, a Base Policy (Transformer-based) is trained to predict not just the next pose, but the expected interaction force.

1. Force as the Ultimate OOD Detector

Instead of complex Bayesian uncertainty or KL-divergence (which are often noisy), the authors use a simple physical intuition: if the force the robot feels matches the force it expected, it’s in-distribution. If there is a spike in discrepancy, it's failed/jammed, and it pings the human.

2. Residual Trajectory Editing

When a human corrects the robot, we don't just throw the data into a buffer. TER-DAgger uses optimization-based editing. It finds a segment of the original trajectory and "warps" it to meet the human's starting point smoothly. This creates four types of training data:

• Pre-editing: Tells the robot "do nothing" when things are fine.
• Transition: Teaches the robot how to steer toward a corrected path.
• Demonstration: The actual human correction.
• Post-editing: Ensuring the robot stays on the new, successful path.

Experiments: Real-World Robustness

The system was tested on notoriously difficult tasks: USB-A, Two-pin, and Three-pin plug insertions.

Significant Findings:

• Success Rates: TER-DAgger achieved 96% on 2-pin plugs in the real world, while standard ACT only managed 64%. On the harder 3-pin plug, it hit 82%, whereas other DAgger variants fell below 25%.
• Detection Precision: The force-based detector had 98.8% precision. This means almost zero "false alarms," allowing a single expert to manage a fleet of robots effectively.

Deep Insight: Why Residuals?

The genius of this work lies in the Residual Policy. Instead of the model trying to learn the whole task from scratch (Retraining) or overwriting its knowledge (Finetuning), the Residual Policy acts as a "corrective layer" on top of the base. This preserves the base policy's global navigation skills while specializing the residual layers on the "tactile nuances" of the insertion phase.

Conclusion & Future Work

TER-DAgger proves that compliance and force-awareness are not just "nice-to-haves" but fundamental requirements for industrial IL. By treating human intervention as a trajectory optimization problem rather than a simple data augmentation problem, they have created a much more stable bridge for learning.

Limitations: The threshold for force detection currently requires manual task-specific tuning. Future iterations could benefit from an adaptive thresholding mechanism that learns "normal" force profiles across different materials.