AI
Metro-WM: Long-Horizon Latent Planning with Realisable Sub-Goals
Publication: cs.AI updates on arXiv.org
Author: Royson Lee, Fady Rezk, Titouan Parcollet, Timothy Hospedales, Cristina Cornelio
Source published: 29/09/2026, 04:00:00 UTC
Briefing created: 29/09/2026, 05:20:53 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 05:18:24 UTC.
Briefings
Metro-WM is a hierarchical planning framework that replaces ungrounded latent sub-goals with real states from prior experience, significantly improving long-horizon success rates by 37.33 percentage points compared to the next best hierarchical approach. The system constructs a graph from observed frames to enable robust path recovery from execution errors, achieving up to 10.9 times faster planning while requiring 13-56 times less offline compute. Key findings include shorter paths than offline demonstrations and superior performance under sparse data conditions, though the specific macro planner used for evaluation is not named.
Key points
- Metro-WM uses a graph constructed from observed frames to connect states across different episodes, enabling robust path recovery when the micro planner drifts off course.
- The framework achieves superior long-horizon success rates of up to 37.33 percentage points over the next best hierarchical approach, outperforming an oracle relying on the query's own demonstration.
Context and limitations
The evaluation relies on a leading state-of-the-art macro planner whose specific architecture is not named. Results are based on a single preprint version (arXiv:2609.31868v1) and may not reflect final model performance. Offline compute requirements are estimated but not measured directly.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 05:20:53 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeide6gs5kueywgffxpaheu5uy2z7xhpxhretr2kact4xctzps3lp2e