Metro-WM: Long-Horizon Latent Planning with Realisable Sub-Goals

Royson Lee; Fady Rezk; Titouan Parcollet; Timothy Hospedales; Cristina Cornelio

cs.AISep 25, 2026

Metro-WM: Long-Horizon Latent Planning with Realisable Sub-Goals

Authors: Royson Lee, Fady Rezk, Titouan Parcollet, Timothy Hospedales, Cristina Cornelio

Organizations: Samsung AI, Cambridge, UK · University of Edinburgh, UK

Abstract

Model-predictive control with Joint-Embedding Predictive Architectures (JEPAs) provides a strong zero-shot goal-reaching planner, but it is only effective over short planning horizons. Hierarchical extensions attempt to bridge this gap by learning a macro planner to predict intermediate latent sub-goals to guide the micro planner. In this work, we demonstrate that unconstrained latent sub-goal prediction is fundamentally flawed. A rigorous evaluation reveals that a leading state-of-the-art macro planner routinely emits physically unrealisable sub-goals. To resolve this, we introduce Metro-WM, a hierarchical framework that issues sub-goals by retrieving genuine states from prior experience rather than generating ungrounded latent vectors. Specifically, Metro-WM constructs a graph whose vertices are observed frames from offline expert demonstrations or random-action trajectories, allowing frames from different episodes to be connected and stitched into routes to the goal. Planning over the full graph also makes the system highly robust to execution errors: if the micro planner drifts off course, Metro-WM instantly finds a new optimal path from the current state. Our experiments show that Metro-WM achieves superior long-horizon success rates of up to 37.33 percentage points over the next best hierarchical approach while being up to 10.9 times faster, requiring both 13-56 times less offline compute and fewer tuned hyperparameters. Additional analysis reveals that Metro-WM finds shorter paths than the offline demonstrations, outperforms an oracle relying on the query's own demonstration, and maintains robust performance under extremely sparse dataset conditions.

Figures & tables

Figure 1: One cycle of Metro -WM. Rather than generating sub-goals, Metro -WM retrieves them from previously seen episodes. The macro planner searches the graph whose vertices represent observed frames encoded as latent states, with $e_{1}\dots e_{5}$ in the figure indicating the specific episodes containing these frames. Directed temporal edges (solid grey) connect consecutive frames within each episode, while symmetric bridge edges (dashed grey) link visually similar frames across episodes (example endpoints in Appendix Figure 7 ). All edge costs are expressed in frame units: Temporal edges have unit cost, while bridge-edge costs are converted into frame-equivalent costs derived from empirical dataset quantiles. Given a goal $z_{\mathrm{goal}}$ , we use Dijkstra’s algorithm to search backward from $z_{\mathrm{goal}}$ to find the lowest-cost route from every graph vertex. At each planning cycle, Metro -WM add the current observation $z_{\mathrm{obs}}^{t}$ to the graph, and identifies the shortest route (blue) to $z_{\mathrm{goal}}$ (which may stitch together multiple episodes). The furthest frame along this route that is within the micro planning horizon is then selected as the next sub-goal $z_{\mathrm{sub}}^{t}$ . After execution (solid orange), planning resumes from the reached state $z_{\mathrm{obs}}^{t+1}$ (see Appendix Figure 4 ).

Figure 2: Decoder-free subgoal analysis. Top row: closest realizable states to the non-realisable sub-goals emitted by Hi-LeWM-C macro planner (Eq. 3 ), drawn by the environment’s own renderer. Bottom row: Our Metro -WM sub-goals, which are real states retrieved from the training dataset. Micro actuation - constrained to realisable states - will not progress toward the final goal when pursuing sub-goals issued by Hi-LeWM-C, but will progress when using Metro -WM sub-goals.

Figure 3: Performance and cost comparison between Metro -WM and baselines across varying budgets. Four benchmarks

\times

three goal distances (planning horizons)

d

	Metro -WM preprocessing (1 GPU)			Hi-LeWM-C training (4 GPUs)
benchmark	encode	index	total	grad. steps	wall	GPU-hours
PushT	$17.1$	$6.0$	$\mathbf{23.1}$ min	$\phantom{0}76{,}215$	$4.01$ h	$\mathbf{16.0}$
Cube	$34.2$	$4.7$	$\mathbf{38.9}$ min	$118{,}125$	$9.09$ h	$\mathbf{36.4}$
TwoRoom	$\phantom{0}7.6$	$1.4$	$\mathbf{\phantom{0}9.0}$ min	$\phantom{00}8{,}370$	$0.50$ h	$\mathbf{\phantom{0}2.0}$
Reacher	$24.4$	$4.7$	$\mathbf{29.1}$ min	$118{,}125$	$6.58$ h	$\mathbf{26.3}$

Table 1: One-time offline cost per benchmark. Metro -WM’s “encode” is one forward pass of the frozen encoder over every recorded frame; “index” is the exact nearest-neighbour table plus the edge-cost computation. Hi-LeWM-C’s training figures are wall clock at the original authors’ recorded recipe; GPU-hours are wall clock

\times

devices.

	Metro -LeWM		Metro -PLDM
benchmark	re-entry	w/o re-entry	re-entry	w/o re-entry
PushT	$\mathbf{62.67\pm 4.62}$	$49.33\pm 3.06$	$\mathbf{44.00\pm 5.29}$	$39.33\pm 6.43$
Cube	$\mathbf{92.67\pm 3.06}$	$84.00\pm 0.00$	$\mathbf{89.33\pm 5.03}$	$88.67\pm 2.31$
TwoRoom	$\mathbf{100.00\pm 0.00}$	$76.67\pm 4.16$	$\mathbf{100.00\pm 0.00}$	$61.33\pm 5.77$
Reacher	$\mathbf{100.00\pm 0.00}$	$99.33\pm 1.15$	$\mathbf{100.00\pm 0.00}$	$98.00\pm 0.00$

Table 2: Success (%) with and without graph re-entry,

d=75

		Success (%)		Arrival Ratio
Approach	Benchmark	Metro -WM	oracle	Metro -WM	oracle
Metro -LeWM	PushT	$62.67\pm 4.62$	$\mathbf{64.00\pm 2.00}$	$\mathbf{2.71}$	$2.89$
	Cube	$\mathbf{92.67\pm 3.06}$	$72.00\pm 4.00$	$\mathbf{0.44}$	$1.32$
	TwoRoom	$\mathbf{100.00\pm 0.00}$	$96.67\pm 2.31$	$\mathbf{0.93}$	$1.65$
	Reacher	$\mathbf{100.00\pm 0.00}$	$93.33\pm 3.06$	$\mathbf{0.65}$	$2.38$
Metro -PLDM	PushT	$44.00\pm 5.29$	$\mathbf{72.00\pm 4.00}$	$\mathbf{3.08}$	$3.26$

Table 3: Comparison with a privileged oracle , where the sub-goals come from the query’s episode own trajectory at inference. Arrival ratio is the median, over the successful episodes, of frames-to-success divided by the expert demonstration’s own first-hitting time;

<1

means arriving sooner than the demonstration .

Approach	Corpus Size	Cube	PushT	Reacher	TwoRoom
Metro -LeWM	$100\%$	$92.67\pm 3.06$	$62.67\pm 4.62$	$100.00\pm 0.00$	$100.00\pm 0.00$
	$10\%$	$90.00\pm 5.29$	$46.00\pm 2.00$	$98.67\pm 1.15$	$100.00\pm 0.00$
	$3\%$	$79.33\pm 1.15$	$38.00\pm 2.00$	$100.00\pm 0.00$	$100.00\pm 0.00$
	$1\%$	$72.00\pm 2.00$	$31.33\pm 4.16$	$100.00\pm 0.00$	$100.00\pm 0.00$
Metro -PLDM	$100\%$	$89.33\pm 5.03$	$44.00\pm 5.29$	$100.00\pm 0.00$	$100.00\pm 0.00$
	$10\%$	$90.67\pm 1.15$	$29.33\pm 5.03$	$100.00\pm 0.00$	$100.00\pm 0.00$

Table 4: Success (%) against corpus size,

d=75

1\%

is 187 episodes on PushT and 100 elsewhere.

Figure 4: Metro -WM’s graph re-entry. This follows Fig. 1 : after the micro planner stops at

z_{\mathrm{obs}}^{t+1}

, the next macro step re-enters the graph there following the optimal shortest route from

z_{\mathrm{obs}}^{t+1}

z_{\mathrm{goal}}

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Top row: closest realisable Hi-LeWM-C’s sub-goals from its macro planner found by Eq. 3 and drawn by the environment’s own renderer. Bottom row: our method’s, METRO-WM, sub-goals which are retrieved real states of other demonstrations in the dataset.

Figure 6: Latent distance stops resolving time within a few dozen frames. Median

\lVert z_{t}-z_{t-d}\rVert

between two frames of the same demonstration

d

frames apart (band: interquartile range), divided by the median distance between frames of two different , randomly drawn demonstrations. A curve at

1

cannot tell a state

d

frames away from a random frame of the corpus.

Figure 7: A bridge, in pixels, on each benchmark. Each row is part of one route computed by Metro -LeWM for one query of the evaluation set. The ends are the query’s $z_{\mathrm{obs}}$ and $z_{\mathrm{goal}}$ and every panel between them is a recorded frame the route traverses, and $\cdots$ marks the stretches not shown. The route runs along one demonstration (blue), crosses a bridge, and continues along a different one (red). The two heavily outlined frames are that bridge’s endpoints, which comes from different demonstrations in the dataset.

Approach	Success (%)	$\Delta$
LeWM (ft)	$15.33\pm 3.06$	$+1.33$
Metro -LeWM	$72.00\pm 3.58$	$+9.33$
Oracle	$81.33\pm 3.06$	$+17.33$

Appendix

Table 5: We fine-tune PushT LeWM, dubbed LeWM (ft) to get a stronger

d=25

model and then evaluate it at

d=75

\Delta

is the gain over performance of the released LeWM weights, on the same

50

episodes.

benchmark	no. of episodes/demonstrations	no. of frames
PushT	$18{,}685$	$2{,}261{,}996$
Cube	$10{,}000$	$1{,}970{,}000$
TwoRoom	$10{,}000$	$880{,}809$
Reacher	$10{,}000$	$1{,}970{,}000$

Appendix

Table 6: Total number of demonstrations and frames for each benchmark in our experiments

$d$	macro: samples $\times$ iterations, top- $k$	micro: samples $\times$ iterations, top- $k$
25	$900\times 20$ , $10$	$300\times 30$ , $150$
50	$1500\times 40$ , $10$	$900\times 30$ , $150$
75	$1200\times 60$ , $10$	$1200\times 30$ , $150$

Appendix

Table 7: CEM hyperparameters of the hierarchical baselines, variance scale

1.0

at both levels. They come from the authors’ released matrix, which covers PushT and Cube ; we apply the same rows on TwoRoom and Reacher .

Long-Horizon Planning

Long-horizon planning studies how agents coordinate actions over extended sequences to achieve goals amid changing states and uncertainty.

See all 58 papers

Apr 3, 2026

Hierarchical Planning with Latent World Models

World models are a promising path to zero-shot embodied control through planning. However, existing world model planners struggle on long-horizon, multi-stage tasks: prediction errors compound and naive search is exponential in the planning horizon. Hierarchy mitigates both by decomposing tasks into shorter, tractable subproblems; yet prior hierarchical approaches either amortize control into task-specific policies (hierarchical RL) or assume low-dimensional states and known dynamics (classical hierarchical MPC). We present Hierarchical Planning with Latent World Models (HWM), an architecture and planning paradigm for hierarchical model predictive control directly on visual world models trained solely via next-latent prediction. HWM learns world models at multiple temporal scales within a shared latent space, so predictions from the long horizon model serve as subgoals for the short horizon model via latent matching, without task-specific rewards or hierarchical policies. To keep long-horizon search tractable, HWM learns an action encoder that compresses primitive action chunks into latent macro-actions. On real-world Franka manipulation, HWM solves pick-and-place from a single goal image at 70% success vs. 0% for single-level planning. Across simulated push manipulation and maze navigation, HWM consistently improves performance on long-horizon tasks while requiring up to 3x less planning compute.

Wancong Zhang, Basile Terver, Artem Zholus +7

FAIR at Meta · New York University · Mila - Québec AI Institute +1

Hierarchical Planning Hierarchical Robot Control Long-Horizon Robotic Manipulation Model Predictive Control Latent World Models

Ask PDF

Oct 19, 2026

Learning Goal-Conditioned Subgoals through Frozen Hierarchical World Models

Hierarchical world models plan over long horizons by passing predicted states from a coarse model to a lower-level controller. Choosing these states can be expensive: the upper-level planner repeatedly differentiates through candidate macro-action trajectories. We replace this search with a lightweight goal-conditioned network trained through the frozen high-level predictor. The network supplies a conditioning vector, while the predictor generates the next latent waypoint. Training uses hindsight goals and state-prediction losses on offline trajectories, without search-solution targets or macro-action labels. Inference retains an autoregressive high-level rollout, but requires no high-level candidate optimization. We evaluate this approach on the shared-latent HWM checkpoints distributed with H-JEPA, using three backbone seeds and three head initializations per backbone. With a common evaluation protocol, mean success is 70.4% on Cube, 48.7% on Push-T, 100.0% on FourRoom, and 71.1% on Visual AntMaze. The method improves over the corresponding HWM search baseline on three environments and has the same mean on Push-T. Dedicated measurements on an NVIDIA L40S give 1.28-4.39x end-to-end speedups for batched evaluation. Ablations identify state-adaptive conditioning as a consistent contributor, while predictor-mediated subgoals and unrestricted conditions have environment-dependent benefits.

Xiang Kaibo, Liu Qingfeng

Department of Industrial and Systems Engineering, Hosei University

World Model-Based Planning World Models Latent-Space Planning Efficient Inference

Ask PDF

May 26, 2026

Latent Goal Prediction from Language for Model-Based Planning

Planning with world models is bottlenecked by compounding prediction errors and the difficulty of defining optimizable goals. Visual targets provide precise local gradients but poor distant guidance, while language is flexible yet limited by noisy cross-modal alignment or dependence on large generative models unsuited for the high-sampling nature of model-based planning. To address these challenges, we introduce Latent Goal Prediction from Language (LAGO), a framework that predicts both sequences of intermediate goal states from language instructions and action-conditioned rollouts, all within the same latent space. Rather than optimizing toward a single global objective, LAGO dynamically decomposes instructions into explicitly predicted, locally tractable latent subgoals. By updating these subgoals online and using a soft minimum trajectory cost during planning, LAGO enables an agent to follow coherent latent trajectories over long horizons. Evaluation across multiple environments planning horizons shows that LAGO avoids the sharp degradation of prior methods. By achieving robust and precise long-horizon planning purely from language, LAGO bridges the precision of visual goals with the flexibility of text-guided control.

Samuel Barbeau, Simon Roy, Giovanni Beltrame +2

1École de Technologie Supérieure, Montréal · 2International Laboratory on Learning Systems (ILLS) · 3Polytechnique Montréal +4

World Model-Based Planning Long-Horizon Planning Model-Based Planning Latent World Models

Ask PDF

Abstract

Long-Horizon Planning

Explore similar work

Hierarchical Planning with Latent World Models

Learning Goal-Conditioned Subgoals through Frozen Hierarchical World Models

Latent Goal Prediction from Language for Model-Based Planning