Concat IDM
An auxiliary training method that feeds the concatenation of latent states from two consecutive time steps into an inverse dynamics model. Using existing action labels in offline trajectories, it updates a shared encoder to achieve action predictability, thereby improving representation learning in world models without requiring additional reward or action annotations.
Timeline (last 90 days)
Concat IDM concatenates the latent states of two consecutive time steps and feeds them into the inverse dynamics model, using the action labels already present in the offline trajectories to update the shared encoder, without changing the planner structure at test time.
On the Push-T task, using 24 different appearances for three independent training runs, Concat IDM raised the planning success rate from 9.1% to 77.2%.
Compared with the improvement under the single-appearance condition, the paired improvement of Concat IDM under the multi-appearance condition was 52.9 percentage points higher.
All Facts (3)
On the Push-T task, using 24 different appearances for three independent training runs, Concat IDM raised the planning success rate from 9.1% to 77.2%.
50%UnverifiedCompared with the improvement under the single-appearance condition, the paired improvement of Concat IDM under the multi-appearance condition was 52.9 percentage points higher.
50%UnverifiedConcat IDM concatenates the latent states of two consecutive time steps and feeds them into the inverse dynamics model, using the action labels already present in the offline trajectories to update the shared encoder, without changing the planner structure at test time.
50%