New phase-conditioned AI boosts robot T-shirt hanging from 56% to 87%

Standard imitation learning policies like Action Chunking with Transformers (ACT) rely on a Markovian assumption, causing state aliasing when visually similar observations require different actions and preventing autonomous failure recovery. To address this, a team of researchers led by Dayuan Chen and Kai Tang introduces a phase-conditioned, force-aware framework that uses a closed-loop hierarchical architecture.

A FiLM-conditioned ACT encoder modulates feature extraction based on the current task phase, allowing a single unified policy to produce phase-specific behaviors while sharing action dynamics. A multi-modal phase predictor fusing visual, force, and pose feedback estimates the phase in real time, detecting contact failures invisible to vision alone and triggering autonomous recovery. A hybrid impedance controller enables compliant execution, paired with a haptic teleoperation interface for force-aware data collection. Ablation studies show FiLM-based modulation significantly outperforms baselines, and t-SNE analysis confirms well-separated phase-specific features. Validated on hanging and removing a T-shirt with dual arms, the closed-loop system improved hanging success from 56% to 87% through autonomous error recovery. Code and videos are available.