Scripted Demonstrations

Demonstration data generated automatically by a hand-written rule-based program controlling the robot, instead of a human.

Scripted demonstrations are produced by a hand-written rule-based program (a scripted policy) that controls the robot through a task automatically, with the process recorded as demonstration data. The script typically reads privileged information directly in simulation, such as an object's exact pose (precise state that isn't available in real deployment), and computes a trajectory using preset waypoints or a motion planner. The advantage is low cost, unlimited generation, and clean, consistent motion; the disadvantage is a narrow range of motion patterns, very different from the diverse, hesitant style of real human demonstrations — complex tasks are hard to script at all, and precise state is hard to obtain on a real robot. It's commonly used for simulation benchmarks, algorithm debugging, and large-scale synthetic data. For the same number of demonstrations, a policy trained on scripted data often reaches a higher success rate than one trained on human teleoperation data, simply because scripted data lacks the multimodality and noise of human motion — so methods validated only on scripted data should be viewed cautiously.

ExampleThe ACT paper recorded 50 scripted demonstrations and 50 human-teleoperated demonstrations for the same block-transfer task in ALOHA simulation; ACT reached 97% success on the scripted data, dropping to 82% on the human data.

Also called

Scripted Policy Demonstrations, Programmatic Demonstrations

Related

Demonstration Data · Simulation Data · Privileged Information · Action Multimodality · MimicGen · ALOHA Sim (Transfer Cube / Insertion)

Sources

Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT, arXiv 2304.13705)
tonyzhaozh/act (GitHub)