Most robotic manipulation benchmarks assume target objects stay still and that a policy's inference is instantaneous. That hides a real problem: by the time an action is executed, the observation may be stale. The new LIBERO-Dynamic benchmark makes these dynamics explicit by extending all 40 LIBERO tasks so objects move under contact friction along three trajectory families and speed tiers, while the simulated world advances according to calibrated, model-specific inference latency.
Beyond evaluation, the benchmark supports training: it replays verified static demonstrations and compensates robot actions online for target displacement, producing 9,000 quality-filtered dynamic trajectories. The authors evaluated six pretrained policies on 1,800 dynamic conditions per configuration, compared execution strategies and action horizons with fixed checkpoints, and fine-tuned three architectures using dynamic and mixed data.
Results show that faster motion degrades every policy, but not uniformly. Execution timing and action horizon can change the behavior of the same policy, and asynchronous replanning only benefits from frequent feedback when the action queue avoids underflow. Dynamic demonstrations improved each architecture under at least one recipe, with gains depending on temporal conditioning, data composition, and training budget. A small-scale physical study found task-dependent execution-strategy differences consistent with simulation.