AgenticTask expands RoboCasa to 16,450 saved task instances, with generation and execution sharing one task record and repair raising the saving yield
Synopsis
Training embodied policies needs many usable interaction tasks, yet specifying and validating each task instance is costly. AgenticTask pairs a task-generation agent with a task-execution agent around one shared hierarchical task record: the generator reuses interaction structure under human intent and revises configurations from validation feedback, while the executor uses the same entity bindings and geometric records to select robot skills and check completion. On RoboCasa the pipeline built a collection of 16,450 saved task instances, and diagnostic-guided repair raised the saving yield above resampling at the cost of more time and model calls per saved instance.
Interpretation
Expanding a task collection no longer requires hand-specifying each instance; agents generate and screen usable tasks under interaction requirements set by human intent. Earlier automated generation largely translated high-level requests into task descriptions or code, while candidate tasks still had to be instantiated, checked against the environment and revised when incompatible. The AgenticTask collection built on RoboCasa contains 16,450 saved task instances, produced at a measured throughput of saved instances per elapsed hour, with a stated fraction of all construction attempts passing the configured physical checks.
The generation and execution agents share a hierarchical task representation whose three levels, semantic requirements, object and part bindings, and geometric scene graphs, tie task intent to concrete simulator entities. The hierarchy separates reusable interaction requirements from instance-specific bindings and geometric state, so instructions, initial conditions and completion criteria keep referring to the same physical targets when objects or scenes change. The generation agent repairs candidates within the permitted configuration space using construction and validation feedback, and the execution agent selects robot-supported skills from the same bindings and geometry and independently checks the rollout against the original completion criteria.
Diagnostic-guided repair and resampling are distinct ways to recover candidates, and repair trades higher construction cost for a higher saving yield. Under fixed task requirements and checking conditions, diagnostic repair raised the saving yield above matched resampling while time per saved instance and reported model calls per save both increased. The comparison reports each strategy's saved fraction, seconds per saved instance and calls per save, placing repair's benefit in output rather than construction speed.
Perspective
The demonstrated scope is task-instance construction and execution testing in the RoboCasa simulator: the collection holds 16,450 saved instances, execution results come from rollouts that restore each task's initial state under its original goals and constraints, and successful cases span object placement, articulated control, door closure and sequential arrangement.
The collection inherits the coverage and accuracy of available assets, category and part annotations, and simulator interfaces; physical and robot-compatibility checks establish configured conditions only within their search budgets, and complete solvability needs further evidence; execution outcomes depend on the model, skill library, robot and retry budget, leaving task-versus-solver attribution unresolved for some failed trials; deployment under real-world perception and transfer across embodiments remain untested.
