# VLAW: Policy and World Model Improve Together

VLAW: Policy and World Model Improve Together

Full article: https://haiguangboy.com/posts/vlaw

## Core Method

Real execution calibrates imagination, imagined experience improves the policy
VLAW starts from π0.5 and Ctrl-World, alternately updating the policy and the world model.

↳ ① Real-robot execution calibrates the world model. Each category is first warm-started with 25 expert demonstrations; each round collects 50 real trajectories per category, and both successes and failures are used to fine-tune the world model, while mixing in the original DROID data to prevent overfitting.

↳ ② Generate more practice inside the world model. Each round generates 500 synthetic trajectories per category, and the policy keeps outputting actions based on predicted observations, forming a closed-loop imagination. Qwen3-VL-4B is fine-tuned with the real success/failure labels from the first round, then filtered by a "yes" probability threshold; not all 500 are successful.

↳ ③ The policy learns only from successes, then returns to real-robot execution. Real successful trajectories and synthetic successful trajectories are jointly used for flow-matching training, and failed samples are given zero weight in the policy loss. The updated policy then collects real data again, entering the next round. Here, failed experience mainly helps the policy indirectly by calibrating the world model.

## Key Results

· Compared with fine-tuning only on expert trajectories, after adding online data, on 50 interaction segments, the world model reduced the number of failures predicted as successes from 11 to 1; it increased the number of successes predicted as failures from 2 to 4, so not all errors decreased.
· Under the same real-robot interaction budget, VLAW achieves a larger success-rate improvement than Filtered BC and DSRL.
· In the circle-drawing task, reducing synthetic trajectories from 500 to 250, or removing real successful trajectories, both lead to performance degradation.

## Comparison with Related Lines

- Opposite line · [NVIDIA's action flow: a new standard for robot actions](https://haiguangboy.com/posts/hydra-0-action-flow) `hydra_0_action_flow_for_generalist_world_modeling_and_control_2026_08`: still relies on real-robot interaction, manual reset, and success labels contradicts the redefinition of the modeling object: not reproducing task-specific behavior, but modeling the consequences of robot motion
- Same line · [PointWorld: point flow unifies state and action](https://haiguangboy.com/posts/pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation) `pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation_2026_08`: PointWorld: point flow unifies state and action
- Same line · [RL should not learn from scratch, it should do post-training](https://haiguangboy.com/posts/rl-100) `rl_100_performant_robotic_manipulation_with_real_world_reinforcement_learning_2026_08`: still relies on real-robot interaction, manual reset, and success labels validates that autonomous reset remains an unsolved practical bottleneck, which the authors themselves list as future work
- Same line · [S1: the GPT-3 emergence moment for robots](https://haiguangboy.com/posts/skild_blog_introducing_s1_in_context_learning_20260817) `skild_blog_introducing_s1_in_context_learning_20260817_2026_08`: S1: the GPT-3 emergence moment for robots
- Same line · [VGEBench: generalizable visual exploration](https://haiguangboy.com/posts/towards_generalizable_visually_grounded_exploration_of_household_devices) `towards_generalizable_visually_grounded_exploration_of_household_devices_2026_09`: VGEBench: generalizable visual exploration
- Same line · [It's not "whether to have a world model", it's "whether what it outputs has structure"](https://haiguangboy.com/posts/robointer15_a_holistic_intermediate_representation_suite_for_embodied_world_mode) `robointer15_a_holistic_intermediate_representation_suite_for_embodied_world_mode_2026_07`: It's not "whether to have a world model", it's "whether what it outputs has structure"

## Boundaries

⚠️ Boundaries
Only five task categories and two rounds of iteration were validated, and manual reset and success annotation are still required. Adding online data simultaneously changes the data volume, state coverage, and composition of failure samples, and the causal contribution of failure samples has not yet been isolated separately.

## Author's Judgment (not part of the paper, a cross-paper synthesis view)

💡 Viewed together with the WAM survey
The survey reminds us: world models pursue prediction accuracy, while policies pursue return, and the two have different objectives. VLAW grounds the division of labor in the data: the world model learns from both successes and failures, while the policy learns from filtered successful experience. They improve together, without requiring both to learn the same set of trajectories.

Looking back half a year later: under the same real-robot budget, after scaling up tasks and iteration rounds, can this alternating improvement continue to yield gains.