In addition to the open-loop rollouts on the project page, we run the policy closed-loop on the real arm-hand system at 1x speed (30 Hz). Videos are played at 1x. Background regions are blurred for anonymity.
Both policies combine motion tracking with goal-reaching training (50/50 TrackEnv/GenEnv). They differ only in how goals are sampled in the goal-reaching environments.
Left — Reference-plus-noise (AdaDex): goals are sampled by perturbing future poses from the demonstration.
Right — Random goals: object goals are sampled independently of the demonstrated trajectory.
Each policy receives the type of goals used during its training. The left follows perturbed demonstration goals; the right follows independently sampled random goals. This shows the behavior each training objective encourages.
Both policies receive the same goal sequence, with matched Stage-2 iteration counts. These examples examine how the two training approaches retain and extend the demonstrated manipulation skills.
The same goal sequence asks both policies to perform two consecutive in-hand rotations.
The same noisy-reference goal sequence is given to both policies.
The same noisy-reference goal sequence is given to both policies.
These are selected qualitative examples, not a random sample of all benchmark outcomes. Some paired videos hold the final frame when one rollout ends earlier.