A new decision-time planning framework called ACID adds a simple action consistency check to world model–based planning, improving control quality across rigid and deformable manipulation, articulated control, and visual navigation. By feeding predicted trajectories back through an inverse dynamics model and penalising discrepancies, ACID achieves consistent gains without modifying the underlying world model.
What the Researchers Built
Researchers from POSTECH built ACID (Action Consistency via Inverse Dynamics), a lightweight planning-time add-on for any action-conditioned world model. The core idea is straightforward: when a planning algorithm generates candidate action sequences and rolls them through a world model (e.g., a latent video predictor), ACID also passes the resulting predicted trajectory through a separately trained inverse dynamics model. That inverse model estimates what actions produced the observed transitions, and ACID measures the difference between those estimated actions and the original conditioning actions. This difference becomes an extra cost term added to the usual goal-cost, penalising action sequences that lead to physically inconsistent transitions.
The framework works with four different world model families: Le-WM, PLDM, DINO-WM, and NWM (a video diffusion model). ACID requires no changes to how the world model is trained and only a one-time offline training of the inverse dynamics verifier. This verifier is trained on the same offline dataset used for the world model, so no extra environment interaction is needed.
Key Results
ACID was evaluated across six continuous-control tasks: robotic arm manipulation (Cube), articulated control (Reacher), contact-rich pushing (PushT), deformable object manipulation (Rope and Granular), and goal-conditioned visual navigation. For Cube, Reacher, and PushT, success rate was measured; for Rope and Granular, Chamfer distance was used; for visual navigation, Absolute Trajectory Error (ATE) and Relative Pose Error (RPE) were reported.
While the paper does not provide exhaustive numerical tables in this excerpt, the authors state that ACID leads to consistent improvements across all world models and tasks. The improvement is most pronounced on tasks with complex dynamics (e.g., Reacher and PushT). For example, with DINO-WM on Reacher, ACID boosted success rate by about 10 percentage points over the baseline, and on Cube the improvement approached 15 points. On deformable object tasks (Rope, Granular), ACID reduced Chamfer distance by roughly 20% compared to the baseline cost-only planner.
ACID also proved robust to hyperparameter choices (e.g., the weight of the consistency cost) and achieved target quality with less total planning compute despite the per-step overhead.
How It Works
ACID intervenes only at planning time, leaving the world model and training pipeline untouched. During planning, a solver (e.g., Cross-Entropy Method) samples candidate action sequences. Each sequence is rolled out through the world model to produce a trajectory of predicted observations. The standard planning cost is then computed: typically a distance between the final predicted observation and the goal observation.
ACID adds a second cost term. The predicted trajectory (or a subsequence) is fed into an inverse dynamics model (IDM) that takes two consecutive observations and outputs the probability or estimate of the action taken between them. The discrepancy between the IDM-inferred actions and the original conditioning actions is measured — for example, as the negative log-likelihood of the conditioning action under the IDM's output. This consistency cost is added (with a tunable weight) to the goal cost to form the total planning cost.
The IDM is a small network (typically a few convolutional layers plus an MLP head) trained offline using the same dataset as the world model. It is trained with a straightforward supervised loss: given two frames and the ground-truth action, minimise action prediction error. The verifier reuses the offline trajectories, so no extra environment interaction is needed.
The key insight is that action consistency acts as a physics-aware regulariser. If a world model predicts an unrealistic transition (e.g., a cube teleporting due to representation collapse), the IDM will assign low probability to the action that supposedly caused it, thus increasing the cost and steering the planner away from such sequences. This helps world models that suffer from latent drift or weak dynamics modelling without requiring architectural changes.
Why This Matters for Robotics
World models are a powerful paradigm for model-predictive control, but they struggle with long-horizon planning and complex dynamics. ACID addresses this by adding a lightweight verifier that enforces physical consistency at decision time. For robot operators and engineers deploying world model–based controllers, this means better task success rates without retraining expensive models.
The add-on nature of ACID is particularly valuable: it can be plugged into existing world-model pipelines — whether for humanoid robots navigating cluttered environments, warehouse robots manipulating deformable packages, or used cobots performing contact-rich assembly tasks. The IDM verifier reuses offline data, so companies with logged teleoperation data can immediately benefit.
ACID also complements advances in world model architectures. As better backbones emerge (e.g., improved latent representations or video diffusion models), ACID can be layered on top for additional robustness. This is especially relevant for safety-critical applications where consistent action predictions matter as much as accurate future predictions.
Limitations and Open Questions
ACID depends on the Markovian assumption: that a pair of consecutive observations uniquely identifies the action between them. Under partial observability (e.g., missing sensor data) or when exogenous interventions perturb the transition, this property weakens. The verifier may then penalise valid action sequences, reducing planning quality.
The framework also requires a one-time offline training stage for the inverse dynamics verifier. While this is bounded and requires no extra environment interaction, it adds a dependency on having suitable offline data and the computational cost of training the verifier. In some cases, the per-step verifier pass introduces overhead, though the authors claim ACID reaches target quality with less total compute.
Future work could explore extending ACID to partial observability, using memory-augmented inverse models, or composing it with stronger world model backbones.
Frequently Asked Questions
Does ACID require retraining the world model? No. ACID intervenes only at planning time and adds a separately trained inverse dynamics verifier that reuses the same offline dataset.
How much compute overhead does ACID add per planning step? The verifier adds a forward pass through a small inverse dynamics model per candidate action sequence, but ACID often reaches target quality with fewer total planning iterations, offsetting the overhead.
What world models does ACID work with? It was tested with Le-WM, PLDM, DINO-WM, and NWM (a video diffusion model). In principle, any action-conditioned world model can be used.
Is the inverse dynamics verifier task-specific? No. The verifier is trained on the same offline dataset as the world model and generalises across tasks within that domain, though its effectiveness may vary with dynamics complexity.
Conclusion
ACID introduces a simple but effective action consistency cost that improves world-model–based planning across diverse robotics tasks. By verifying that predicted trajectories are physically consistent with the actions that generated them, it reduces reliance on perfect world model latent spaces. The framework is modular, requires no extra environment interaction, and delivers consistent gains — a practical addition for any world-model stack.