Paper ↗
● Decision-time planning · World models

ACID: Action Consistency via Inverse Dynamics for Planning with World Models

A predicted trajectory can look like it reaches the goal while the real one drifts away. ACID checks every predicted step by asking an inverse dynamics model which action it implies.

¹POSTECH · ²KAIST · †Corresponding authors

Does the imagined plan actually happen?

Standard planning scores a candidate only by how close its predicted final state is to the goal. Nothing checks whether each predicted step could actually be produced by its action.

Below, DINO-WM plans the same episode twice. In each pair, the left frame is the real environment and the right frame is what the world model imagined, with the real object overlaid in magenta. Step through the plan and watch where they split.

Start o₀ Original cg ✗ Fail Ours cg + waca ✓ Success Goal og
OriginalThe imagined rope lies right on the goal line, but the real rope (magenta) stays where it was. The goal cost is satisfied by a future that can't happen.
OursImagined and real coincide at every step, and the real rope reaches the goal.
WhyACID rejects candidates whose predicted steps the inverse dynamics model can't explain with the commanded action.
Figure 3: qualitative comparison on Rope and Granular with DINO-WM

Close the loop with Cycle Action Consistency

The true outcome of an action is unknown until it is executed, but the action itself is known. ACID moves the check into action space.

① Roll forward

The world model Fθ predicts the next latent ẑt+1 from ẑt and the conditioning action at.

② Infer backward

An inverse dynamics model Gϕ reads the predicted transition (ẑt, ẑt+1) and infers the action ât that explains it. A flow-matching decoder; one Euler step is enough.

③ Compare

If the step is realizable, ât recovers at. Otherwise the residual ‖at − ât‖² exposes the world model's error. Under locally linear dynamics it bounds that error from both sides (Prop. 1).

④ Re-rank

CEM minimizes cg + waca, with a scale-invariant weight wa = λ·σg/σa recomputed every iteration. The world model is never retrained.

ẑt ẑt+1 Fθ at Gϕ ât ‖a−â‖²

The full pipeline, step by step

Fig. 1, animated in the order of Algorithm 1. Each loop starts with the whole figure, then highlights the modules running on the current line; after J CEM iterations the plan is executed and the loop starts again.

ACID overall pipeline (Fig. 1)
    CEM iteration j = 1 / 3

    Consistent gains on every world model

    Three JEPA-style latent world models, five simulated tasks. Only the planning cost changes; training and the CEM planner stay fixed.

    Original (goal cost cg)Ours (cg + waca)Mean ± std over three seeds (Table 1). Hover a bar for values.
    Success rate (%) ↑ higher is better
    Chamfer distance ↓ lower is better

    Original vs. ours, side by side

    One card per task. Each Original/Ours pair is the same episode and world model, played in sync. Swipe or use the arrows.

    1 / 8

    On a real robot arm

    Policy-guided planning on a PiPER arm with a single RGB camera. A Diffusion Policy proposes 20 action chunks and the planning cost picks one. Same policy, same world model, same candidates; only the cost differs.

    One episode per method, played at 2×. All three rank the same 20 policy samples; only the planning cost differs.

    Less planning compute

    ACID reaches the baseline's full-budget quality with far less search. On Reacher, 30 CEM samples already beat the baseline's 300. With DINO-WM, ACID reaches the baseline's final Chamfer distance in half the planning steps on Granular, and after the first step on Rope. The verifier does add about 12–15% per CEM iteration, but the samples and planning steps ACID saves in reaching the target more than compensate for that overhead.

    Fewer planning steps · DINO-WM

    Mean Chamfer distance over planning steps (lower is better). Orange dashed line: where Original ends after 10 steps.

    BibTeX