Long-VLA: Unleashing Long-Horizon Capability of Vision Language Action Model for Robot Manipulation
Phase-aware masking improved long-horizon performance across multiple base architectures.
The model outperformed prior approaches on L-CALVIN and reported real-world demonstrations.
Unscripted tasks with longer chains, automatic phase labels, recovery after subtask failure, interventions per hour, and cycle-time economics.
Phase-aware masking improved long-horizon performance across multiple base architectures.
pp. 8-9, Scalability and ConclusionThe model outperformed prior approaches on L-CALVIN and reported real-world demonstrations.
Abstract and p. 8, ConclusionTraining-phase decomposition still requires manual work.
The evaluated task and sequence-length scope is limited.
The method reduces initial-state mismatch but does not solve execution failures that accumulate later.
Long-horizon capability is improving, but the evidence is not a fleet-level reliability or cost result.
Thematic