Temporal Policy: History-Initialized Action Generation for Robotic Learning from Demonstration

Researchers introduced Temporal Policy, a generative control framework that seeds action generation with recent robot history.

By replacing an uninformative Gaussian prior, it reduces transport cost and inference latency for action generation.

The method matched strong baselines on simulation benchmarks and a physical Barrett WAM teleoperation setup.

It addresses a key bottleneck in generative robot control by enabling faster closed-loop operation without sacrificing multimodal behavior or task success.