PRIME estimates precise progress from imprecise demonstrations

The PRIME method learns precise progress estimation by aggregating experience across trajectories.

PRIME is a novel progress estimator. It can learn precise progress estimation from imprecise demonstrations.

Existing methods assume all actions in a demonstration contribute to success. This assumption breaks down in high-precision manipulation where geometric errors lead to failure.

These methods can interpret human mistakes as meaningful progress. This can corrupt progress learning because they lack a global prior regarding task evaluation criteria.

PRIME constructs a weighted graph over observed states. It connects states across trajectories based on spatial proximity and visual consistency.

Edge weights encode spatial distances as transition costs. A shortest-path search approximates the geodesic distance to success.

Changes in this distance distinguish advancement, stagnation, and setbacks. In charger insertion, PRIME yielded up to a 5×5\times improvement in policy success rate over SFT baselines.

Key facts

  • PRIME distinguishes advancement, stagnation, and setbacks without assuming monotonic progress over time.
  • PRIME is effective across diverse tasks in simulation and the real world.
  • The advantage of PRIME is more pronounced in higher-precision scenarios.

All sections

Edition