Goal Force: Teaching Video Models To Accomplish Physics-Conditioned Goals

TL;DR

Goal Force reframes video-conditioned planning by allowing users to specify a desired physics outcome as a goal force and training a diffusion-based video model to infer the antecedent causal chain. By introducing a multi-channel physics control signal and curriculum-trained synthetic data, the model learns to act as an implicit neural physics simulator, propagating forces through time and enabling zero-shot generalization to tool use and complex multi-object interactions. The approach achieves superior adherence to physics-based goals compared to text-only baselines and prior force-prompting methods, while preserving visual realism and enabling diverse, physically plausible plans. This work advances interactive, physics-aware world models that can plan actions to achieve specified dynamic outcomes without relying on external simulators at inference.

Abstract

Recent advancements in video generation have enabled the development of ``world models'' capable of simulating potential futures for robotics and planning. However, specifying precise goals for these models remains a challenge; text instructions are often too abstract to capture physical nuances, while target images are frequently infeasible to specify for dynamic tasks. To address this, we introduce Goal Force, a novel framework that allows users to define goals via explicit force vectors and intermediate dynamics, mirroring how humans conceptualize physical tasks. We train a video generation model on a curated dataset of synthetic causal primitives-such as elastic collisions and falling dominos-teaching it to propagate forces through time and space. Despite being trained on simple physics data, our model exhibits remarkable zero-shot generalization to complex, real-world scenarios, including tool manipulation and multi-object causal chains. Our results suggest that by grounding video generation in fundamental physical interactions, models can emerge as implicit neural physics simulators, enabling precise, physics-aware planning without reliance on external engines. We release all datasets, code, model weights, and interactive video demos at our project page.

Figures (6)

• Figure 1: Given a force-conditioned task, goal force enables video models to generate the antecedent action to accomplish the task.
• Figure 2: Goal Force: A user provides an input image and a goal force, and the model generates a video containing a force that locally causes the goal force. Our model generalizes to diverse objects and interactions and enables visual planning, respecting the physical properties of the objects and their environments.
• Figure 3: Force prompt and goal force prompt result in different behaviors. With a direct force applied to the red block (top), the effect is directly materialized (i.e. the block falls over). The force in this case is encoded in the red channel of the control signal as a moving Gaussian blob. In contrast, with a goal force applied to the red block (bottom), the model must find the antecedent motion to achieve the goal force (i.e. the pendulum swings to knock over the block). The force in this case is encoded in the green channel of the control signal as a moving Gaussian blob. We visualize the control signal overlaid on top of the video via alpha blending.
• Figure 4: In prior methods (right), the user provides a force, and the model directly applies the force to the target object. In our method (left), the user provides a goal force, and the model generates the causes that achieve the desired effect on the target object. The top three methods (PhysGen liu2024physgen, PhysDreamer zhang2024physdreamer, and Force Prompting gillman2025forcepromptingvideogeneration) all accept forces as conditioning; the fourth method, Tora zhang2024tora, accepts trajectories rather than forces, so we condition on an acceptable trajectory.
• Figure 5: Given a goal force prompt, the model chooses the physically correct way to execute it. Top: even though there exist multiple plausible initiators, the model correctly selects the white ball as the initiator to achieve the desired force on the target. Bottom: With multiple plausible rubber ducks that could initiate the force, the model selects the initiator that is not blocked by a physical barrier.