LAGO turns language instructions into latent subgoals for planning agents
Researchers behind the LAGO world model say one predictor both forecasts dynamics and converts instructions into latent subgoals, reporting gains on long-horizon navigation and manipulation tasks.
In plain words
LAGO is a planning system that lets an agent follow a written instruction by breaking it into intermediate goals inside its own internal representation, then choosing actions that reach them. The authors report it beat several image-goal planners and a vision-language reward model in their tests, especially on long navigation tasks.
What was announced
A team introducing LAGO, short for Latent Goal Prediction from Language, describes a hierarchical world model built on the idea that agents can plan by imagining outcomes in a learned representation rather than raw images. The paper argues that visual targets give precise local feedback but weak guidance over long distances, while language is flexible yet often suffers from noisy alignment between modalities or requires separate large generative models. LAGO addresses this with a single predictor that handles two jobs: forecasting how the world changes under actions, and translating an instruction into a chain of intermediate latent subgoals. Both are trained together with one regression objective over a shared latent space.
At each planning step, the system derives subgoals from the instruction and searches for actions using an alignment cost that rewards getting near subgoals without forcing a fixed route. The subgoals are recomputed as the agent moves, so one long instruction becomes a series of shorter, manageable objectives. In tests across navigation and manipulation settings, the authors report that LAGO beat flat and hierarchical image-goal planners, more than doubling the success rate of a goal-conditioned behavior cloning policy on long navigation with curved, hazard-constrained paths. They also say it outperformed vision-language reward models, including one planning with ground-truth dynamics. These results come from the paper's own abstract and have not been independently verified.
What it does not show
- Evidence is limited to the paper's abstract; no full-text results, environment details, or external evaluation are available.
- Performance claims, including the reported doubling of success rate, are the authors' own and lack independent confirmation.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.