Trajectory Phase Labeling
Cross-source consensus on Trajectory Phase Labeling from 1 sources and 5 claims.
1 sources · 5 claims
How it works
Risks & contraindications
Highlighted claims
- Each trajectory chunk is assigned to one of five semantic phases — active-grip, pre-grasp, release-ramp, approach, and tail — using a deterministic gripper-state-based heuristic. — Learn Where Outcomes Diverge: Efficient VLA RL via Probabilistic Chunk Masking
- Phase labels are applied identically to successful and failed trajectories to ensure per-phase comparisons are like-for-like. — Learn Where Outcomes Diverge: Efficient VLA RL via Probabilistic Chunk Masking
- Active-grip and pre-grasp phases consistently dominate PCM's chunk allocation throughout training because they carry the highest success-failure action variance. — Learn Where Outcomes Diverge: Efficient VLA RL via Probabilistic Chunk Masking
- Tail-phase chunks are assigned only floor-probability selection because they represent downstream consequences of failure rather than causally critical moments. — Learn Where Outcomes Diverge: Efficient VLA RL via Probabilistic Chunk Masking
- Gripper-state phase labeling is specific to single-gripper manipulation and has not been empirically validated for dexterous hands or locomotion settings. — Learn Where Outcomes Diverge: Efficient VLA RL via Probabilistic Chunk Masking