Lazy-MOSAIC speeds up long-horizon robot manipulation planning by using a learned model to predict what robot skills will do, then checking candidate plans with a higher-fidelity simulator before trusting them. Across 160 trials, it solved 104 tasks versus 81 for the original MOSAIC planner, and ran up to 6.6 times faster on settings both planners solved—without letting model mistakes validate an incorrect plan.
What did the researchers build?
Lazy-MOSAIC is a planning system for manipulation tasks that require a robot to perform a sequence of skills. A skill is a reusable action, such as moving an object or carrying out another manipulation step. Planning with skills can make long tasks easier to manage, but the planner still needs to work out which sequence is likely to reach the goal—and what state the robot will be in after each step.
The system combines a learned world model with a search strategy called lazy planning. The world model predicts the effects of skills. Because those predictions are cheaper to query than the high-fidelity model used to check a plan, Lazy-MOSAIC uses them to guide its search rather than treating every prediction as confirmed.
That distinction changes what the planner can defer. In ordinary lazy search, a planner can postpone checking an action on a known state. Here, the state produced by a skill may also be uncertain. Lazy-MOSAIC therefore treats both the planned steps and the states between them as hypotheses. If a prediction proves wrong, the planner revisits the affected part of the search rather than assuming the rest of the sequence remains valid.
The goal is to spend expensive simulation effort where it is useful, while keeping a final plan grounded in the high-fidelity model.

What results did Lazy-MOSAIC achieve?
Across eight benchmark settings, Lazy-MOSAIC matched or exceeded MOSAIC’s success rate. It solved 104 of 160 trials, compared with 81 for MOSAIC. Among trials both planners solved, Lazy-MOSAIC also took less time on the reported easy, medium, and hard settings, with comparable solution lengths.
| Reported comparison | MOSAIC | Lazy-MOSAIC |
|---|---|---|
| Successful trials across 160 trials | 81 | 104 |
| Easy-setting average time on successful trials | 124 s | 43 s |
| Medium-setting average time on successful trials | 331 s | 83 s |
| Hard-setting average time on successful trials | 479 s | 216 s |
| Task 3, hard-setting successes | 0 | 4/20 |
The largest reported per-setting speedup was 6.6 times, on Task 3 Medium. Task 3 Hard remained difficult: Lazy-MOSAIC solved 4 of 20 trials, while the version without uncertainty handling solved 1 and the other methods solved none.
Other lazy-planning baselines solved far fewer problems as clutter increased. Across the medium and hard settings, Lazy Options solved 9 trials and Lazy Roadmap solved 13, compared with 64 for Lazy-MOSAIC. Giving those baselines access to the learned model reduced simulator queries, but did not produce similarly high success rates. The results point to the importance of how skills are selected and organized, not just how cheaply their effects can be predicted.
How does Lazy-MOSAIC work?
A planner searches for a sequence of skills that takes the robot from its starting situation to a goal. In a skill-centric planner, each step has two linked pieces: the skill to execute and the state expected after it. For example, a plan might predict that one manipulation skill changes the scene in a way that makes a later skill possible. If that predicted change is wrong, later steps that depend on it could be wrong too.
Lazy-MOSAIC uses a learned skill-effect world model to estimate these outcomes cheaply. Rather than immediately checking every proposed skill and every resulting state with the high-fidelity model, the search can first use the learned predictions to explore promising possibilities. This is the “lazy” part: expensive evaluation is deferred until it is needed.
Deferral requires careful handling because an unverified state is not a reliable foundation for the rest of a plan. Lazy-MOSAIC treats predicted states as hypotheses, alongside the skill steps that connect them. When a prediction fails high-fidelity checking, the planner must account for the affected downstream search—rather than merely marking one action as bad while leaving its dependent states untouched.
Before a plan is accepted, certification checks it against the high-fidelity model. That check is the safety net for planning correctness: the learned model can influence which options the search considers and how quickly it finds candidates, but its prediction errors do not make an uncertified plan valid. The paper’s comparison with a non-certified variant tests why this matters: faster prediction alone does not guarantee that the robot reaches its goal.
This design separates two jobs: the learned model helps direct search, while high-fidelity checking establishes whether a candidate plan works under the evaluation model.

Why does this matter for robotics?
Robots often need to plan several linked actions, and evaluating every possible sequence can be costly. That becomes a practical bottleneck when a task has many candidate skills, when object arrangements are cluttered, or when each simulation check takes significant time. Lazy-MOSAIC targets that bottleneck by using inexpensive predictions to focus the search, while retaining a more reliable check before a plan is accepted.
The benchmark results also highlight a point for teams choosing planning systems: a cheap model is not enough on its own. The weaker success of other methods, even when the learned model reduced simulator queries, suggests that search strategy and skill organization are important parts of the overall system. For operations teams, a faster planner is useful only if it also produces plans that satisfy the task.
The reported evaluation is in simulated manipulation tasks, not a field deployment, so it does not establish how the approach performs on a particular commercial robot. Still, buyers and engineers assessing a platform can consider whether it supports skill-based planning and simulation-based validation. Teams comparing platforms can browse humanoid robots on Robot Overflow or explore warehouse robots as examples of robot categories where multi-step manipulation and planning can matter.
What are the limitations and open questions?
The results establish gains on the evaluated benchmark, not across all robots, objects, or real-world environments. The reported evidence comes from simulation trials; it does not show that the same success rates or speedups carry over to hardware, where sensing errors and unmodeled physical effects can change outcomes.
Task 3 Hard also shows that difficult cases remain difficult: Lazy-MOSAIC solved only 4 of 20 trials there. The paper excerpt does not provide enough detail to assess sensitivity to different learned-model training data, model quality, or the cost of certification in a physical deployment.
Finally, certification protects plan validity relative to the high-fidelity model, not necessarily the real world. A key open question is how to build and validate that model so it remains a dependable reference when the robot, objects, or operating conditions change.
Frequently Asked Questions
What problem does Lazy-MOSAIC solve? It speeds up planning for multi-step manipulation tasks by reducing how often the planner needs to evaluate options with a high-fidelity model.
What is a skill-effect world model? It is a learned model that predicts the state changes caused by robot skills, allowing the planner to explore candidate plans at lower cost.
Does the learned model decide whether a plan is valid? No. It guides search, while certification checks a candidate plan against the high-fidelity model before acceptance.
How did Lazy-MOSAIC compare with MOSAIC? It solved 104 of 160 trials versus 81 for MOSAIC, and was faster on the reported settings both planners solved.
Conclusion
Lazy-MOSAIC combines learned predictions with delayed, high-fidelity checking to make skill-based manipulation planning faster without relying on the learned model as the final judge. Its benchmark results are promising, while the hardest tasks and the gap between simulation and real-world operation remain important tests.
