Project Page
Reflection-iteration agents for controllable, physically plausible video generation.
A small red ball rolls down from the higher step on the right and lands in the paper cup below; the impact then makes the cup lose balance and tip over.
Speed of the red ball should be faster to fall into the cup.
The red ball should be heavier to knock down the cup, and increase the speed a littlt bit.
A bowling ball curves to the right along the lane, hits only the right-side pin, knocks that pin down, and leaves the other pins standing.
The bowling ball's velocity direction should be adjusted further to the right.
The bowling ball's velocity is now deviating too far to the right, so it should be adjusted slightly back to the left.
A red curling stone slides from behind toward a yellow stone at the target, hits it, pushes it away from its original position, and finally stops near the center of the target.
The red curling stone's velocity direction should point straight toward the yellow stone.
The red curling stone should move faster, and its mass should be slightly increased.
A yellow toy stands on a wooden table; a blue ring descends from above but does not fully loop around it, instead getting half caught on the toy body and stopping at an angle.
Increase the ring's initial downward velocity.
Slightly reduce the initial velocity to avoid overshooting the desired contact state.
A small rocket launches from the grass, first flies upward to the right, then gradually curves left, forming a C-shaped arc.
The rocket thrust should be stronger and last longer.
The steering force should be applied slightly earlier.
A snowball flies in an arc toward a wooden fence, hits the fence. Some of it breaks into powder, and bounces back outward.
The snowball collider should be fixed during the early stage to prevent it from breaking apart in flight.
The snowball needs a higher initial velocity.
In a side-scrolling pipe game scene, a small bird flies to the right while bobbing up and down to adjust its height, passing through the gap between each pair of vertical pipes in sequence.
A red toy car drifts sideways while descending the transparent ramp and the car tips over onto its side on the tabletop.
A knife is thrown above a wooden cutting board, flies forward while flipping in the air, and finally lands near the board.
Two oranges slide off the end of a conveyor belt, fall to the floor, collide with a third orange that was resting below, and push it forward.
A huge pumpkin hangs directly above a small wooden stool, falls vertically, lands heavily on it, and crushes the stool.
A scooter rolls across the ground toward a trash can, hits it, loses balance and falls over, while the trash can is also knocked down.
In a snake game, the snake first moves to the right to eat a piece of food, then turns downward and continues moving.
A snowman stands in the snow; the hat on its head is blown off by the wind and falls to one side.
In a stack of books near the table edge, the top three books are pushed off the tabletop and fall, while the bottom book remains on the table.
Only the top book in the stack is pushed sideways; it slides off the stack, flips once in the air, and lands flat on the floor, while the remaining books stay upright in a stack.
After the support plate beneath the transparent ramp is pulled away, the white foam block loses support and falls downward.
A white foam block is above a transparent ramp and gently floats downward through the air, slowly approaching the ramp surface.
In a snake game, the snake first moves to the right, then turns downward and continues moving to eat the fruit.
In a snake game, the snake moves through four successive stages, changing direction step by step and eating the target food.
A snowman stands in the snow; the hat on its head is blown off by the wind and falls to one side.
A snowman is hit by a sideways wind or force.
An umbrella lies beside an open road; strong wind lifts it off the ground, carries it into the air, and makes it drift around.
An umbrella is blown by the wind on an open road; instead of flying high, it tumbles backward along the road surface.
A red toy car starts on a transparent ramp, rolls down the slope, and reaches the flat tabletop.
A red toy car rolls down the transparent ramp, then turns around and changes direction after reaching the bottom.
A small yellow cube rests near a table edge; as its center of mass passes the edge, it slips outward and falls.
A small yellow cube near the table edge receives a sideways force, flips to the left, and falls off the tabletop.
A soccer ball on the grass rolls forward across the field and crosses the goal line.
A soccer ball is kicked from the grass, leaves the ground, flies toward the goal, and lands inside.
A red kite held by a string in the sky sways left and right under the wind.
A red kite in the sky is pulled by the wind and moves along a roughly circular path.
Physically plausible video generation remains challenging for current generative models, especially when the user prompts describe fine-grained object dynamics, force interactions, complex trajectories, and temporally structured motion. Existing physics-grounded video generation methods introduce physics engines to provide explicit simulation priors, but they often rely on one-shot vision-language model prediction to configure physical parameters and motion controls. Such a paradigm is fragile due to the strong coupling among simulation parameters, the limited expressiveness of low-level force or collision-based controls, and the lack of feedback mechanisms for realizing complex and precise physical outcomes.
In this paper, we propose a reflection-iteration agent framework that closes the loop between vision-language model configuration, physics simulation, reflective evaluation, and iterative parameter refinement. Beyond improving the control of coupled physical parameters, our framework enables the agent to progressively realize complex trajectories, multi-stage interactions, and precise event outcomes by treating each physical program as an executable hypothesis. In addition, we design a set of high-level physics-control APIs to support more compositional and complex motion behaviors. Our framework improves prompt alignment, physical plausibility, and generalization in physics-grounded video generation.
Starting from an input image and event prompt, the agent infers a physical control program, executes it in simulation, verifies intermediate artifacts, and refines failed factors before using the accepted rollout as a structured motion prior for video synthesis.
