X-Planner: Event-Structured Task Planning for Embodied Intelligence
X-Planner makes robot task planning explicit by turning long-horizon reasoning into event-structured states.
The paper presents a planning front end for embodied AI systems, aimed at the layer between high-level instructions and executable robot behavior. It uses data from Ego, UMI, and teleoperation sources, with annotations that include takeover moments and human-designed failures for error recognition. The model exposes both an interpretable discrete event interface and a latent planning interface using Staircase Decoding. In offline evaluation, it ranked second among four models on BERTScore-F1 and a judge-based overall score, and the authors report stronger real-robot results than evaluated baselines. HF Daily Papers' note
The paper presents a planning front end for embodied AI systems, aimed at the layer between high-level instructions and executable robot behavior. It uses data from Ego, UMI, and teleoperation sources, with annotations that include takeover moments and human-designed failures for error recognition. The model exposes both an interpretable discrete event interface and a latent planning interface using Staircase Decoding. In offline evaluation, it ranked second among four models on BERTScore-F1 and a judge-based overall score, and the authors report stronger real-robot results than evaluated baselines. HF Daily Papers' note
score 5