AdaDex
A hybrid dexterous imitation framework that trains a shared policy simultaneously in a massively parallel motion-tracking environment and a generalization environment. It combines dense tracking rewards with sparse success rewards to learn dexterous hand skills that possess both human priors and generalization ability from a small number of HOI trajectories.
Timeline (last 90 days)
AdaDex is a hybrid dexterous imitation framework that can train dexterous hand skills with only a single several-second hand-object interaction trajectory.
AdaDex's core mechanism is to share one policy while training in two types of massively parallel environments: motion tracking environments (dense rewards) and generalization environments (sparse rewards).
AdaDex is tolerant of both motion capture data and HOI trajectories reconstructed from in-the-wild videos, reducing reliance on expensive motion capture equipment.
The skills demonstrated by AdaDex cover functional grasping, object repositioning, in-hand reorientation with finger gaiting, and dynamic tossing and catching.
The authors verified AdaDex's cross-simulator transfer and real-world deployment.
The scale of AdaDex's real-world experiments, its sensitivity to reconstruction noise, and its performance in long-horizon tasks still await more details.
All Facts (6)
AdaDex is a hybrid dexterous imitation framework that can train dexterous hand skills with only a single several-second hand-object interaction trajectory.
50%UnverifiedAdaDex's core mechanism is to share one policy while training in two types of massively parallel environments: motion tracking environments (dense rewards) and generalization environments (sparse rewards).
50%UnverifiedAdaDex is tolerant of both motion capture data and HOI trajectories reconstructed from in-the-wild videos, reducing reliance on expensive motion capture equipment.
50%UnverifiedThe skills demonstrated by AdaDex cover functional grasping, object repositioning, in-hand reorientation with finger gaiting, and dynamic tossing and catching.
50%UnverifiedThe authors verified AdaDex's cross-simulator transfer and real-world deployment.
50%UnverifiedThe scale of AdaDex's real-world experiments, its sensitivity to reconstruction noise, and its performance in long-horizon tasks still await more details.
50%