When to Act: Calibrated Confidence for Reliable Human Intention Prediction in Assistive Robotics

TL;DR

This work tackles safe triggering in assistive robotics by using calibrated confidence to decide when to assist. It introduces a lightweight multimodal GRU-based next-action predictor that produces a calibrated confidence $\\hat{c}$ via post-hoc methods such as Temperature Scaling and Isotonic Regression, feeding a safety-driven Act/Hold gate. The approach significantly reduces calibration error (from about $ECE \approx 0.40$ to as low as $0.039$) without sacrificing accuracy, and demonstrates improved act-only precision across varying coverage via the safety threshold $\\tau$ with a bound $\\mathrm{AOP}(\\tau) \ge \tau - \\varepsilon$. This yields a practical, verifiable safety mechanism for real-time assistive control, though prospective user studies and online recalibration are left for future work.

Abstract

Assistive devices must determine both what a user intends to do and how reliable that prediction is before providing support. We introduce a safety-critical triggering framework based on calibrated probabilities for multimodal next-action prediction in Activities of Daily Living. Raw model confidence often fails to reflect true correctness, posing a safety risk. Post-hoc calibration aligns predicted confidence with empirical reliability and reduces miscalibration by about an order of magnitude without affecting accuracy. The calibrated confidence drives a simple ACT/HOLD rule that acts only when reliability is high and withholds assistance otherwise. This turns the confidence threshold into a quantitative safety parameter for assisted actions and enables verifiable behavior in an assistive control loop.

Figures (4)

• Figure 1: Calibrated assistance pipeline. The system (1) senses multimodal context, (2) encodes intention with a multimodal GRU, (3) calibrates scores so that confidence matches empirical accuracy, and (4) uses a hysteretic Act/Hold gate on $\hat{p}$ to trigger reliable assistance.
• Figure 2: Reliability diagram showing how calibration aligns predicted confidence with empirical accuracy. Isotonic regression follows the identity line more closely than the uncalibrated model.
• Figure 3: Calibration performance across methods. Temperature scaling and isotonic regression reduce ECE substantially while leaving Top-1 accuracy unchanged.
• Figure 4: Closed-loop act-only precision as a function of the confidence threshold $\tau$. Calibration turns the threshold into a meaningful safety parameter: the calibrated gate achieves high precision already at mid-range $\tau$, while the uncalibrated scores remain flat until extreme thresholds.