Same world model, two ways to plan
CEM samples hundreds of candidate action sequences, imagines each one with the world model, and refits. HIJP's inverse model proposes the action directly, so a single imagined rollout per replan is enough.
One PushT episode on the extreme tier (goal 75 steps ahead, 100-step budget). The latent trajectories are projected to 2-D with PCA. HIJP and the CEM ablation share the same encoder and world model; only the planner differs. The faint fan shows 200 action sequences sampled from CEM's initial proposal and imagined with the world model. Each replan covers 25 environment steps; CEM evaluates 300 × 30 sequences × 5 steps = 45,000 latent transitions per replan.
Train the inverse model inside the world model
HIJP jointly trains a visual encoder fθ, a latent predictor gφ, and a goal-conditioned inverse head hξ, end-to-end from offline trajectories without rewards. Click a panel to see which modules each loss updates.
Hindsight relabeling: every future frame is a goal
Pick an anchor, then sample a goal offset κ ~ Unif{1, …, W−1−i}. The target is always the immediate action ai, but the goal can sit at any horizon in the window.
Window of W = 16 observations spaced k = 5 environment steps apart (PushT). Click a frame to set the anchor, and click a later frame to set the goal. Fixing the goal to the window's last frame (the "endpoint goal" ablation) drops overall success from 82.5% to 71.7%.
Search-free planning: alternate inverse and predictor
From the current latent and the fixed goal latent, the inverse head proposes an action chunk and the predictor advances the imagined state. That is H = 5 steps, then execute and replan.
Four tasks, four horizons
Goals are placed K steps ahead along an offline trajectory, from easy (K = 10) to extreme (K = 75). Success is judged only by the simulator, never by latent distance. Pick a task and tier to watch HIJP plan; the curve tracks the latent distance to the goal frame by frame.
Normalized latent distance ‖zt − zg‖ / ‖z0 − zg‖ in HIJP's latent space. Click the curve to jump the video. Latent distance is a diagnostic only: in TwoRoom, for example, the agent detours through the doorway and the distance can rise before the goal is reached. These are representative rollouts as in the paper's Appendix E: the longest successful episode on easy, default and hard, and a failure on extreme when available. The OGCube extreme clip is marked visually unsuccessful because its final frame differs from the goal image, although the source rollout is flagged successful by the simulator, which checks only the cube position.
Higher success, a fraction of the planning cost
All methods use the same offline datasets and evaluation protocol: 50 episodes per task–tier combination, success judged by the simulator.
Success rate
HIJP versus eight baselines. Gains are largest on long-horizon manipulation (PushT and OGCube on the hard and extreme tiers).
Data table (Table 1)
Planning time
Total wall-clock time for 50 episodes in one batch, with matched replanning calls. The label shows the speedup.
Data table (Table 2)
Ablations
Remove one ingredient at a time. Mean success over Reacher, PushT and OGCube (TwoRoom is saturated).
Data table (Table 3)
Latent progress toward the goal
Normalized latent distance on the extreme tier (K = 75), mean ± 95% CI over 50 episodes. HIJP and its CEM ablation share one latent space; LeWM uses its own encoder.
Smoother actions
Action jerk, the mean ‖at − 2at−1 + at−2‖ (lower is smoother). Because HIJP and its CEM ablation share the same representation, the difference comes from the planner.