X-Planner: event-structured task planning front-end for long-horizon robot manipulation
Howard Lu · hf · 2026-09-24
X-Planner is a planning front-end for embodied AI addressing implicit plan structure in VLA systems and coarse supervision in CoT planners.
- Data: combines Ego, UMI, and teleoperation under hierarchical granularity with source-dependent annotation depth; takeover-time annotations and human-designed failures supervise ongoing error recognition.
- Model: a shared VLM backbone exposes a discrete interface emitting interpretable event states and a latent interface relaying continuous CoT states across staggered Transformer depths via Staircase Decoding, with a frozen latent-to-text reconstruction objective as semantic anchor.
- Results: ranked second of four models on BERTScore-F1 and judge-based Overall in offline two-step planning; outperformed baselines in real-robot experiments.
More from Embodied
- Meta's Michael Bronstein says AI-directed hardware in the lab is already happening — mmbronstein · 2026-09-24
- Baidu's Apollo Go robotaxis have tested on London streets for over a month — Baidu_Inc · 2026-09-24
- European humanoid robot startup Vesoma unveils Alpha platform, touting self-evolving agent — CyberRobooo · 2026-09-24
- Qualcomm powers all of Meta's new product announcements, capping a strong week — BenBajarin · 2026-09-24
- Manufacturing jobs: outsourcing vs tech — and why factories create the designers — lemire · 2026-09-24
- IROS 2026 Robotic Origami Challenge pits dexterous robot hands against origami masters — micoolcho · 2026-09-24